I work at OpenAI and I can try to explain our logic.
In general, LLM performance is incredibly multidimensional. There's no guarantee of a fixed ordering where the model that is best at math is also the model that's best at jokes. Despite that complexity, much of the variation in performance is explained by just a single variable: model size. In general, models that are bigger will (a) be more performant / reliable, (b) cost more per token, and (c) have better recall of esoteric facts. As you note, our models are ordered:
Luna < Terra < Sol < Astra
Even today, I'd recommend this as a general starting point for picking models. If you care about quality only, pick Astra. If you're a bit price sensitive, pick Sol. If you're very price sensitive, pick Terra or Luna.
Of course, model size doesn't capture all variation. Models that are trained later might have new data or new techniques or new architectures that earlier models lack. You can see this with models from all providers - Sonnet 5.5 outperforms Fable 5.1 in ways, and Gemini 3.8 Flash outperforms 3.1 Pro in ways.
In the case of GPT-6.1 Terra, I'd say Terra is probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things. For API developers, it's a nice midrange option if you want something cheaper than Sol but a lot smarter than Luna (especially for vision / computer use / communication). For subscribers, it's not a bombshell release - if you're hitting usage limits, maybe it helps you go a bit further, and if you want clearer communication or faster computer use, it can be a nice option to select.
If you have suggestions on how to name our products better, I'm all ears. Our intention is to make things clear and simple, even if we miss the mark sometimes.
Unfortunately, a lot of the complexity is somewhat irreducible and due to to the actual complexity of the products. Scaling model size is different from scaling reasoning effort which is different from scaling multi-agent count. So if we just swept all of those under a single name of "more IQ", it would make the naming simpler but the experience much worse, as important details would be obfuscated from end users.
Do you then recommend Sol for technical documentation?
What's confusing to me in addition to the naming is that there are no published guidelines I can use, as far as I can see. Put yourself in the user's shoes. On what basis should one pick a model for a given task?
I can understand smaller models being dumber and faster. That aligns well with scenarios where you are willing to compromise; e.g. you're classifying content or asking a model to format text into a table.
But when you want "maximum intelligence, cost and time be damned", there's the implication that bigger models are smarter on a linear scale, and every provider's model release naming is currently conflicting with that expectation.
This isn't just a naming problem, but a communication problem. If I knew that Astra was the best at coding but a bit bad at writing, and Sol was almost as smart but better at technical writing, then that's the kind of thing I expect to be documented.
One possible way: A simple matrix with horizontal columns like "writing", "coding", "math", rows are the models, and each box is a number of stars from 1-5.
Instead, this sort of thing becomes lore, shared by word of mouth.
> Do you then recommend Sol for technical documentation?
I don't know, to be honest. Technical documentation is such a multidimensional skill, and I also don't know your budget. I'd imagine Terra might have the most readable prose, but Astra might be more likely to catch bugs or mistakes to highlight. If the documentation is about a more esoteric language or situation, Astra might have a greater edge than if it's a common language or situation. If you have a long list of documentation guidelines, that might also give more of an edge to Astra. I wouldn't want to make a recommendation without doing a lot of testing, and I haven't done a lot of testing.
Some food for thought, if someone from oAI can't make a direct recommendation for something as direct as "Technical documentation is such a multidimensional skill" - how could we know assess what model to use?
To elaborate, I can certainly ask each model to write a doc about something I know well and tell you which output I like better.
But to assess whether my preference on my problem is going to generalize across other people, other programming languages, other documentation types/formats, etc., and then assess statistical significance is a much broader question that I personally don't have the data to answer. I'd rather make no recommendation than a poorly informed recommendation.
It’s overkill most of the time, but frankly all these models are “cheap”, and I’d rather waste compute than brain cycles trying to discern which dials are best for each individual as-hoc query.
> Put yourself in the user's shoes. On what basis should one pick a model for a given task?
Create a realistic eval suite for your use case using your actual data that you manually label, with easier, more complex, rarer samples, and run each model you are considering against it and decide based on the usual confusion matrix, cost, latency, or whatever metrics matter to you. There's really no other way to know – and you might be surprised that cheaper models can often perform your task(s) relatively well at a fraction of the cost.
> Create a realistic eval suite for your use case using your actual data that you manually label, with easier, more complex, rarer samples, and run each model you are considering against it and decide based on the usual confusion matrix, cost, latency, or whatever metrics matter to you.
I’d imagine there’s only a tiny percentage of users that can or want to do this.
I recently experimented with some image generation models and I was surprised at how much context/prompting is needed to get something that doesn't look Lovecraftian. There was seemingly no downside to adding more and more detail up to the limit. Prose sits between code and images in terms of fuzziness and I wonder if promptmaxxing is the winning strategy for technical documentation.
I've been thinking that fine-tuning a model to write differently shouldn't be that hard (tedious, perhaps) and if I had the resources I would consider doing it myself.
Humans cannot be in the loop for autonomous work. I have agents review code and design/architecture documents, write alert messages based on rules, write analyses (e.g. outage incidents), and many other things. "Having a person write words" would mean not using agents, defeating the whole point.
I do not use agents to write on behalf of myself. I write my won technical write-ups and my own slide presentations. A lot of people don't; and it shows. There's much that agents can do better than humans now, but not everything.
> I'm all ears. Our intention to make things clear and simple, even if we miss the mark sometimes.
Thanks. I think the very response you wrote here is a better form of marketing for the software engineers than what openAI puts out there. When everything is groundbreaking and everything tops charts, nothing is.
I suspect a lot of target demographic (developers) prefer simpler and honest communication about not just the models high points but low points too. With the linked Terra post, if I saw 1 chart that clearly highlighted the usecases for which Terra is _worse_ than higher tier models, I would be more convinced overall.
Everything trying to occupy the top spot in my brain's limited "top model" representation space is not helpful.
While I generally agree that it'd be nice to just have "here is the Pareto optimal model for your use case. Here is the specifics we optimized each model for" so users don't have to do benchmarking and research as much, they're getting much better at explaining. It's basically what the commenter above said.
The website linked starts out with:
> We’re introducing GPT-6.1 Terra, a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing. A significant step up from GPT‑5.6 Terra, it excels at the communication and computer-based tasks that fill the workday, from drafting proposals and interpreting technical drawings to filling out forms and editing videos.
If you go father down, it has graphs and details with specifics of this, including:
> We measured GPT-6.1 Terra’s computer-use capabilities... At medium reasoning effort, GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra at about one-third the cost. It even performs on par with GPT-6 Astra at only one-tenth its price per token.
and
> On MMMU-Pro... It even nearly matches the scores of our most capable frontier model, GPT-6 Astra, and GPT-6.1 Sol at much lower cost.
The 4-tier naming is honestly as clear as you can get. I have no idea what people are so confused about.
Terra and Luna are dirt cheap and good enough for 50% (or more) of the typical junior dev's workload, and great for routine work. Sol and Astra have so much written about them at this point that it should be obvious what those are good at. The large/gigantic models for things that need full intelligence and slightly more branching thinking.
The obnoxious/confusing thing is that models are released individually rather than as a family.
Luna < Terra < Sol < Astra, fine. But does Terra 6.1 beat Sol 6? Does Fable 5.1 beat Opus 5.5? If I build something around this Terra 6.1 release, then when Luna 6.2 is released in a month, should I swap?
I understand the confusion, but isn't this kind of similar to asking does the iPhone 14 Max beat the iPhone 16 Mini? At least with phones there are hard specs we can use to compare, but in a lot of cases the answer will be "it depends"
At least there is a simple system - the words 'Mini' and 'Max' are intended to refer to the size of the item (that one is miniature, and the other is their maximum size).
If they were called iPhone 14 Luna and iPhone 16 Sol then this is less clear.
Or on the flipside, if the models were called GPT 6.1 Mini and GPT 6.1 Max, nobody would be arguing about the names being hard to grok. The words 'mini' or 'small' are more descriptive than 'Luna' or 'Terra'.
The problem with that is then you would have GPT 6.1 Max with max thinking which is different from GPT 6.1 Small with max thinking.
If you look back the naming problem started around the point that OpenAI started doing reasoning models (o1, 4o) where they had to describe this fork in model behavior that let a smaller model think longer and be smarter.
You can come up with naming schemes that don't clash, but there's something to be said for this approach where the thinking level is like you describe and the model name requires you to learn that Astra is the largest model.
Sure, in a hypothetical world where the latest Max phone is two generations older than the latest Mini phone. In which case now you have to evaluate not only screen size but ALSO specs and component age and software support etc. How badly do you need the bigger screen? Etc
If you are comparing the iPhone 16 Max to the iPhone 16 Mini, everything is a lot easier.
Run your own evals then if you have this much of a need to "correctly" choose. That'll make it clear.
I'll just stick to mostly Sol on low/medium/high and start experimenting with Terra tomorrow onward to see what it does noticeably better than Luna/Deepseek flash.
Any pareto frontier model, especially from the leading labs is guarnateed to be good enough for whatever your expectations are at this point. A quick few day test should be enough for anyone to make their own conclusions.
Or I'm going to give up and not use any. Consumers aren't perfectly rational. This isn't like toothpaste, where, despite being overwhelmed at the store, I still need to brush my teeth so ahhhhhh fuckit and just pick a toothpaste out of the 30 available. The alternative is I keep using Google search and think AI is stupid. Or just use Facebook's thing.
What would you want to see instead? Not sure about you but my needs are specific enough that I don't think OAI has evals for them, and I wouldn't want to give them that data even if they wanted to run it for me.
- Understanding how effort and class interact (Is a high effort terra above a low effort Sol? Are larger models mostly always better regardless of effort?)
- understanding when (if ever) a new model tends to be better than an older but higher class one. Is an old astra better than a new terra?
If increasing effort and increasing model class produce improved results in a different way, I miss making it explicit in which way. In other words: What kind of issue warrants going up in effort? what kind of issue warrants going up in model class? What kind of issue is unlikely to be solved by either?
OpenAI itself cannot answer these questions. A prospective price-sensitive (or speed-sensitive) user has to experiment to find the cheapest (or fastest) model that is sufficient for each task.
Since internally you probably have a fairly good intuition for what these models do well, what about assembling a benchmark suite (with multiple dimensions) that produces results that show that intuition, but in a 'somewhat objective' way?
That could solve the 'more IQ' problem, because you'd actually have different dimensions (e.g. for 'raw intelligence', technical writing, emotional understanding, etc) that you could use to communicate the differences in models better.
(Of course people would instantly start complaining that the benchmarks numbers are probably fake etc, but it is what it is.)
Internal evals are still running but GPT-6.1 Terra is an AWESOME release for anyone using the API in agentic production applications where cost matters.
Personally, I pretty much exclusively use claude for coding work (sorry!). Our application is running almost 100% on OpenAI API (except for some fallback models). It is the best API on the market. The models are excellent particularly for any application that is cost sensitive.
What makes this even more confusing is reasoning effort.
You have Luna, Terra, Sol and Astra, each with 4+ reasoning levels. That's ~20 options to choose from.
On a repetitive task like customer support or something, where you have decent benchmarks, you can plot cost versus performance, but you can't really do that on coding.
I actually do have a recommendation, which is to do something like a new mode (i know i know):
Auto xlow
Auto low
Auto med
Auto high
Auto xhigh
Auto max
With an optional flash mode.
Then as you release models, you pick the model and thinking level that you believe is pareto optimal for cost/performance.
So this auto updates each release, like Terra was probably off this list entirely, but maybe this would appear now in Auto medium or something like that.
I know from looking at charts that each model seems to have sweet spots (like I always use this model in this thinking level and if i want to go cheaper i go here). If you picked those for us, it would be interesting. I do think you should keep the model families for those wanting more consistent experiences, but for many devs they may want to just pick auto something and have it update over time. I think nice if it actually just said in parenthesis what it was, something like:
Auto Medium (Terra 6.1 High)
I am not sure if everyone would agree here though. I'm obivously advocating for more naming schemes, and i understand the irony.
"In the case of GPT-6.1 Terra, I'd say Terra is probably better than Astra and Sol at technical communication" -- you're doing the thing here too, underspecifying your own models: 6.1 Terra is better than…which Sol, 6.1, 6, 5.6, just the latter two, all of them? This 2D bluriness is a big part of the confusion.
I like to use OpenAI models for ingesting a customer brief, usually a paragraph or two and then a read of a several customer files (eg spreadsheets and docs etc) and turning it into a system design blueprint. I'm currently using Astra for this phase but it's incredibly expensive relative to the other tasks that come after this. Would you recommend I stick to Astra or can I downgrade to say Sol or Terra for this job?
I'm just worried the system design won't be as good if I choose a lesser model so to speak.
Please can I ask about one parameter you didn't mention here? All of your models have both a reasoning effort setting and also a non-pro vs pro setting. I have some intuitive feel (from experimenting) for the difference a larger vs smaller model makes, and why that's different from adjusting reasoning levels, but I have no real feel for how pro vs non-pro is different in behaviour from adjusting effort.
I think you can think of Pro approximately as running multiple instances of a non-pro model, than giving all their outputs to a final judge model which combines them into a better one.
Appreciate that these things have orthogonal measurements; with that in mind, given that the name signifies size, what does version number signify?
I.e. are all version numbers independent and Model X 7.5 is not necessarily of the same generation or cut off as Model Y 7.5? Or do all of your 6.1 models have something in common - cut off data, generation, baseline technology, approach, architecture, or something else?
These days, version number signifies a few things at once: the order of release, the relationship to other releases, and the magnitude of improvement. So something like 6.1 would indicate it was launched after 5.6, may share some similarities to other 6.1 models in terms of training data / personality (which is why it's not 5.7), and isn't a giant enough leap to merit the headlines and expectations that would come with a 7.
Back in the day we used numbers to also sometimes indicate size, which could make it extra confusing: e.g., we launched GPT-4.1 (an incremental update on GPT-4) after GPT-4.5 (a new model that was a big jump over GPT-4).
It’s probably just unavoidable. Like when Apple release a new MacBook Air first and it is (in some ways) faster than the old MacBook Pro.
I think the planets thing is a bit confusing though. I hadn’t really thought about it being based on the size of the planet. I assumed it was based on how far away they were with terra therefore being the base one.
This kind of heuristic is actually extremely useful - and mostly what I try to develop over a couple weeks as the new models come out. I would love if we could get a fairly standardized take like this on the regular.
Of course we have far too many benchmarks (of questionable utility), but it would be helpful to get OpenAI's take on where it slots in, and how the landscape changes as these get rolled out.
It would be interesting to know which model is internally viewed as being best at jokes, I have a personal humor benchmark and so far models will either steal jokes (from comedians like Maria Bamford), or write something very unfunny.
When I was first learning to have agents interact I had a little script to orchestrate a model telling a joke, a critic collaborator helping develop material and an audience emoji rank them.
A favorite…
Prompt: “Tell a joke about mother in laws”
Gemma2 response: “What’s the best kind of mother in law? A dead one.” 0_o
Straight up - why not just name them Cheapest, Cheaper, Expensive and More Expensive? Cost seems to be the main driver of the intuitive difference between models.
Same reason that Apple doesn't name its series of phones the iPhone cheap, iPhone regular, iPhone semi-expensive, and iPhone expensive. It's poor branding, isn't easily extended to future models, and gets confusing when comparing across model generations, where an iPhone 14 expensive might be cheaper than an iPhone 16 cheap.
I love the current naming scheme and I am excited to see the astronomy trend expand in the future. I hope to see asteroids, binary stars, and solar systems at some point soon, as well as neutron stars, black holes, and eventually galaxies, and clusters of galaxies.
> Models that are trained later might have new data or new techniques or new architectures that earlier models lack.
This seems like one of those cases, where CalVer would be appropriate, and you likely have ideas what you aim for while training, so technical model names (ids?) could include aliases to most things you said. Like
2026q4-xlarge-researcher, 2026q4-medium-humanitarian, 2026q4-medium-coder, 2027q1-small-analyst etc.
> important details would be obfuscated from end users.
I don't expect the vast majority of your users care. I would guess they want a rough proxy for "cost" and will assume cost=>capability are roughly linked in a linear fashion.
it is very possible that the users weighted by spend do, and I'm therefore simply wrong. But, 2c
My guess is nothing, and they're currently just releasing them individually as soon they're ready instead of all at once so that they hit the market more quickly. AI moves quickly enough that a delay of a couple weeks could result in significant lost revenue.
Lower models typically use distilled outputs from higher model to improve .
all training runs have to at some level share the compute capacity ; while they have more than enough to run many large runs concurrently, they don’t have enough compute to run everything concurrently.
> I work at OpenAI and I can try to explain our logic [...] general, models that are bigger will be more performant.
> If you have suggestions on how to name our products better, I'm all ears. Our intention is to make things clear and simple, even if we miss the mark sometimes.
How about naming the models:
"Small", "Medium", "Large" and "Very Large"
Or refer to them by parameter count like other companies (e.g. this is 6.1 5T)
If I have a clothing store, and rather than name the clothes sizes Small, Medium and Large I decide to call them Urgon, Planfar and Fangor I shouldn't be surprised if people think my size naming is confusing or silly. And IMO it's not a great explanation to say "oh it's easy, if you are skinny then you should get Urgon, but then if you are a bit more of a medium build you should get a Planfar, but if you want something a bit baggier than Fangor will really work out for you"
The problem with small, medium, large naming is that they don't last. In a few years our names would have to advance to labels like giant, behemoth, gargantua. This is different than clothing, where we know that in five years humans will be most the same size as today.
As a point of comparison, GPT-3 was launched in a series of small, medium, large, and XL. Back then, XL meant a mere 1.3 billion parameters!
This is kind of like Apple with the iPhone. You need 3 model products max. 3 is easily understood and people can be bothered (if they care) to track the difference. More than 3 is a mix of being redolent of ADHD (viz bad PR) while also being annoying for consumers to track the difference.
I can’t tell you the difference between the 10 different versions of the current iPhone and I can’t you the difference between the 5-6 different models put out by the Anthropics/OpenAIs of the world. At this point I just try to track the flagship and then the cheapest one or worst case I just use the flagship.
I’m sure internally there’s data showing people want choice and options but decision fatigue is of course very real.
Very fair. There's always a tradeoff in the number of products, and it's hard to make everyone happy. My suspicion is that the optimal number is probably a bit higher for developers on our API (who often build their own internal evals) than for individual users (who don't want to waste time re-optimizing every month). We're excited by Terra's uniquely fast computer use, but if it traction is poor, maybe we'll go back to 3. It all depends on the market signal - we only want to ship models people find uniquely useful.
I promise you this would become a giant mess. No one wants to pick GPT-good-at-writing-but-not-in-documents-and-decent-at-instruction-following-but-has-some-known-quirks-to-be-aware-of-p.s.-good-at-knock-knock-jokes-MAY-2027.
Even "writing" as a single concept is very multidimensional. There's fiction writing, nonfiction for experts, nonfiction for learners, deep research, quick lookup Q&A, copyediting, writing in documents / artifacts, instruction following about style, and default quirks (e.g., em dashes, negation, loving the word "tapestry"). Some models are great writers but less diverse, meaning people prefer the first story they read but disprefer the tenth. How does one label such a model?
> No one wants to pick GPT-good-at-writing-but-not-in-documents
But this is exactly what we have! Except it's worse, because the model usage recommendation isn't in the API, it's in some random person's X advice that's no longer current this week.
I understand AI companies wash their hands of this problem because it's hard, and they themselves barely understand the true capability matrix of the models they're selling.
I just don't think some effort to categorize models and expose that to users would be messier than what we have -- model selection driven by influencer voodoo bandwagons and analogies to literal astrology.
> If you have suggestions on how to name our products better, I'm all ears.
For some reason, this is just hilarious to me. Idk if you're being serious or if this is passive aggressive but the thought of OAI actually taking advice from a random HNer and basing their naming scheme off that is peak.
This is one of the last bastions of semi-serious discussion on the internet, and it's full of incredibly high level engineering talent.
Where else would you go? I'm honestly scratching my head thinking of someplace that wouldn't run you out on a rail, but that would give serious feedback. Kernel mailing lists are out, maybe there's something like that which I can't remember off the top of my head, but...
Also, nomenclature is incredibly hard. There are only two hard problems in computer science: cache expiration and naming things.
But regular people are not the consumers of OpenAI's API - they consume "Chatgpt" and don't care about model nomenclature.
The primary users of any API are the designers and engineers of the applications that use it. How many laypeople know what Postgresql is? How many of the applications they use run on it?
We aren't talking about the API. We're talking about the model names. If someone fires up the OpenAI website to ask it a question, nobody knows what the names mean. That's what normal people do and are probably super confused at what is going on.
I don't understand what's so confusing about this.
The lineup of models is still the same.
There are four models: luna, terra, sol, and astra.
Pricing increases for each level. This new 6.1 Terra is 50% cheaper than Sol 6.1.
They are lowering prices for newer releases, so yes if you compare Sol 6.1 to 5.6 Terra it won't look right, but why are you doing that? Look at the same release version for comparable numbers.
Each model improves in successive releases but they maintain pricing and capability coherency within a release version.
Yes, Sol 6.1 is supposed to rival Astra for some work, which is what you'd expect from a later versioned release model against a prior release bigger model.
Assuming they release Astra 6.1 at some point, I expect it'll be better than Sol 6.1, at a higher price.
I don't understand what people are asking for here? No reductions in pricing? Different model names on every version release? Not releasing any models for a version number until they're all ready?
Hey! I think it's confusing because there is a mix of three components:
- Model name
- Version number
- Effort level
All 3 of these are a one-dimensional balance of price/speed, and capability (Loose terms). The confusion is in how these overlap, how they're announced/marketed etc. So you go from 1 dimensions to 3; you can visualize it as a box. For a given task, how do you choose where in this box you select? Which regions are your favorite? Can you map each region of this 3D space to specific tasks? Label them with intuitive names?
Given the operations of the model are opaque, can you see how someone (most people?) might be confused?
But what's the ask? This is the cleanest way I see of them doing it.
I see a ton of criticism and not a single person coming up with anything better, unless it's to take away options, like removing effort level.
What about buying Intel CPUs? A car? Electrical components?
If you're not just using this at a consumer level to ask about cooking recipes (and if that's your use case, you really don't care about API pricing, or any of this, and can just use the default), and are using it professionally, then this seems like a low bar of effort to understand what you're buying.
I do grant you the relationship between model and effort level has an unclear relationship, but there are also tradeoffs in most complex technical products that are not always clear; this is exacerbated by the fact that this product is stochastic.
Why are engineers using this as a professional tool expecting this to be bought like a consumer gadget and then blaming the maker?
“The ask” is give me a clue how to choose the best settings for a specific task. Model releases happen so often that the calculation keeps changing. Maybe even OpenAI doesn’t always know the answer.
This morning I asked GPT which models and settings I should use to inspect my VM dashboard with computer control. It said:
- GPT-6 Astra at Medium
- GPT-6.1 Sol at Medium (alt)
- use Luna only when the route was already known.
But I had already completed the task twice with Luna and it was 5.6! It navigated through a few pages of my dashboard, pulled up the 30-day usage chart, analyzed my CPU and RAM, and explained why I didn’t need an upgrade. I’d give Luna 5.6 an A on this task.
Maybe OpenAI should give us a “Tiptronic” selector: recommend the model and reasoning level based on the prompt, while still letting users override it.
> “The ask” is give me a clue how to choose the best settings for a specific task.
What are you talking about? If it's for some slop you vibe check it. If it's for a serious effort you write an eval. Same as it has been for the last 3 years.
You are showing a selection from some third-party interface that mixes together models from six different suppliers: Mistral, Kimi, OpenAI, xAI, Anthropic and Google, with a bunch of old obsolete models.
Are you presenting this as evidence that OpenAI's naming scheme is bad?
If you don't want to spend the time to know the difference between all these different suppliers and their models, then why are you using whatever this program is that gives you so many choices, most of which are bad? Why is this OpenAI's fault?
The OpenAI employee above says that some smaller models of the same generation are better at certain tasks than larger models, regardless of cost difference.
They should have a guide saying "Terra is best at technical documentation and communication, Astra is best at large software projects and research, Sol is good for development" or something like that.
Effort (as it's used in the context, how many "thinking" tokens to burn) is still a free variable. It's not always shown to the end user but you can set it independently for any of these models.
After that you just have model "strength" and version number. But of course strength is only comparable at equal version number and the two axes are almost parallel.
Anthropic also does this, perhaps to a lesser extent.
It's confusing because the models are released one at a time rather than as a family.
Let's say I have some task that Luna 6.1 cannot do, but Terra 6.1 can. So I build my product around that. In a month when Luna 6.2 is released, then I have to figure out whether it's worth to evaluate and swap or just wait for Terra 6.2.
I think a lot of confusion comes from the discombobulated rollout of these models as well.
They released gpt-6-sol (alongside astra and luna but no terra) and 7 days later released gpt-6.1-sol making gpt-6-sol basically DOA before any real adoption. Everyone knows it wasn’t just better RL/post-training, it was a completely different model underneath. Now they randomly reintroduce terra.
I’m sorry, but if you’re not even a tiny bit befuddled by this then you must be taking whatever OpenAI is having.
Part of the confusion is that it's not linear capabilities. One model is not strictly "better" than the other. They have orthogonal capabilities.
I think the names are more of a reference to price. Luna is always cheaper than terra which is cheaper than Sol which is cheaper than Astra (for the same model number).
It's like Intel i3/i5/i7/i9, or nvidia 50/60/70/80.
It is meant to give a rating inside a generation. Different generations are not necessarily comparable. an i7 is better than an i5, but not if you compare an i7 7xxx with an i5 14xxx.
A nvidia xx80 card is better than a xx60 card, but not if you compare a 1080 to a 5060.
The only difference for OpenAI is that not every generation has every model.
But minor credit to Anthropic, at least they've mostly stuck to their four model names (Haiku, Sonnet, Opus, and Fable). You mostly know where a model is MEANT to be positioned from the name alone. OpenAI cannot keep it consistent from release to release.
But, yes, currently Opus Vs. Fable has the same issue as Sol-6.1 Vs. Astra.
I don’t think opus-5.5 is better than fable-5.1. I think opus is better overall because it’s much faster and cheaper than fable but if speed and cost is not a factor fable it’s probably just as if not slightly better than opus, especially at front-end design (tho the difference is marginal and no one should be really using fable day to day due to this alone).
It's like iPhone 17 being faster than iPhone 16 Pro. Because it's newer and benefits from advances in chip architecture, despite being a lower model tier. Sometimes different iPhone models are released at different times.
Both companies have screwed up their naming royally. Maybe that's intentional - I'm sure it doesn't hurt Anthropic to have people using Fable a bit longer than they need to and racking up a huge bill in the process.
They're always very clear about it when a new model comes out. You have to click away a banner to continue using the service. There's not much they can do to make it any more obvious without it being a major hassle for all the people who understand versioning.
If you daily all these models (and have a modest budget, not some >$400 corporate allowance), then it's pretty obvious that Luna is small and fast, and Sol 6.1 is the big one for jobs that need significantly above average intelligence, comparatively to a human engineer. I use both constantly. Astra is really not relevant for the average dev except to do some moonshot task with a vague verification loop.
Sure Sol 6.1 on low/medium can be considered "cheap" considering its abilities, but it's clearly not meant for the same type of tasks as Luna. You just grow an intuition for when something is solvable with a Luna-sized model. I'm guessing Terra is good for the tasks that you'd typically use Sol 6.1-low for, or for Luna type tasks (or other not-novel problem) where you want to reduce the risks of suboptimal decision making. It's exactly what I'd go with for feature additions where the correct solution is already mostly designed/specced-out.
The way to think of this is that (current) inferencing hardware has natural cut points of efficient scale (assuming some optimal architecture which at this point is almost always an MOE variant), so you always have an incentive to provide 3-4 model sizes (which nicely enough often means diff speed to answer also).
You are also rolling the training of your models by size and the lat est small model can sometimes beat an older model.
For a while we thought you just train the big model first then distill smaller models from it but since we've learnt the importance of on policy RL (basically the way a small model figures out a problem is very diff from a big model so it's best to RL optimize each model distinctly)
You're just not paying attention. If you run copilot in copilot using copilot effort, it performs much better than copilot in copilot using copilot effort. This is clearly explained if you visit copilot.com using copilot and then search for copilot.
If you scale a model down, or switch out attention with linear attention, or introduce KV-compression, etc., you don't really know how it will affect model performance before you've trained the model.
And once you've spent millions training it, you might as well make it available if it's pareto optimal for some tasks. It also gives you more insight seeing how people use it in practice.
Agree, it's confusing. I hear people discussing which model to use for what tasks and my head starts to spin because there are so many properties to account for and it's generally hard to reason about LLMs.
"Speed, quality, price: pick two". I think I'd rather see a choice like that than the current convolution of $-per-token, sub quota, latency, simulated reasoning level, model size, context size, etc, etc.
>But then things go confusing, they released Sol-6.1 which sometimes is superior to Astra, and discontinued Terra which made it look like this
I don't want to rush to their defense here, but, when model families are getting refreshed every 6 weeks, there is going to be significant overlap, particularly with these staggered releases
Wait I always thought terra was the smaller model and the further away you went in space the bigger the model (luna is further away so it’s better than terra)
It's especially confusing because Sol 6.1 is mostly just Astra but 10x slower at 1/10th the price - unless you pay for it to be as fast as Astra - at which point it's the same speed and price...
It's nice that they have Terra which seems to fill a void that's been needed since Sol 6 focused on cost (at the expense of quality), Sol 6.1 got a huge boost in quality, but became unusably slow unless you paid Astra prices - at which point, why not just use Astra?
the only confusing bit was dropping terra for a bit,
Luna (Small), Terra (Medium), Sol (Big), Astra (Bigger)
is perfectly clear, sol beats astra sometimes because Astra 6.1 was scrapped. and note that opus 5.5 beats fable 5.1 most of the time too without a skipped model.
Different classes and different generations. Isn't it roughly the same as what car manufacturers do? Or even Macbooks? 2014 macbook pro has worse performance than 2026 macbook air, this doesn't confuse consumers.
Interesting that both "none" and "low" reasoning levels appeared not to spend any reasoning tokens at all, and in fact low ended up using less tokens and costing less than none here.
"medium" did use 496 reasoning tokens, but that's apparently not enough here to return a visible reasoning trace.
It's neat that we can now compare the Terra without reasoning pelican (cost 0.64 cents) with the equivalently priced Luna with reasoning pelicans (Luna xhigh is 0.40 cents, Luna max is 0.91 cents). I think the Terra low one compares favorably to those, though I'd suggest also comparing them on tasks more representative of real work than generating pelican SVGs!
Can anyone give me the compelling reason for Terra and Sonnet to exist?
Virtually every task I give to either is better served by going up a tier and down a reasoning level, or down a tier and up a reasoning level. I simply don't understand their use case. No testing I've done has shown them to have a niche where there isn't a cheaper-with-equal-results or more-performant-at-same-cost option.
I think for the majority of the audience here who use these models inside of coding agent harnesses w/ subsidized monthly subscriptions, there likely isn't one. Depending on how Terra contributes to the rate limits, it may make sense to use it for some subset of a repeated workflow to get more mileage out of sub.
If you pay API prices for inference in your product stack & have specific use-cases w/ custom evals, these smaller models provide an additional cost/speed lever to adjust. Ultimately, I think the point is to compete on cost for use-cases where Sol on the lowest reasoning is overkill.
If you have a fairly straightforward usecase where Terra performs just as well as Sol, but is faster and cheaper, why not use it?
Yes but you are missing the point. It's the case where Luna is not good enough, and Sol is too much. The problem is that on benchmarks Luna xhigh is typically comparable to Sol low in terms of capability. So where does Terra fit in?
On the API we use a lot of Luna (cost efficient) and sometimes Terra level models. I think we never use a larger model. These are models that our app runtimes depend on.
On subscription for coding and stuff I am basically on Astra Ultra all day long.
You should look at model reasoning level curves on the cost/performance chart on artificial analysis. If you compare the curves for haiku/sonnet/opus I would agree with you that sonnet serves no purpose because it is priced too high. If you look at luna/terra/sol you will see that terra is priced such that it successfully fills a gap between luna and sol. And when you add speed into the equation it adds another dimension where terra makes sense. terra low falls between luna xhigh and max on both cost and intelligence, but it is way faster because it uses way fewer reasoning tokens. So I would say there are workloads where terra makes sense. But from my analysis I don't see any workloads where sonnet makes sense.
Sonnet is way faster than Opus when doing everyday tasks, I think on the subscription plans it's a better option when you need something done faster than Opus' long thinking sessions.
There is a chasm of performance difference between haiku and sonnet ime. Sonnet is cheaper and faster than Opus, so if you're providing a service via api where cost and speed is important, it's a big difference.
1. In Anthropic's benchmarks, Opus's advantage over Sonnet diminishes at high reasoning levels. Sonnet max actually beats Opus max.
2. There are rote or simple tasks where higher reasoning isn't needed and one model vs the other wouldn't even be distinguishable.
3. How many tokens are used during reasoning depends as much on the task as the reasoning level.
4. Sonnet is faster and half the price per token, so Opus will only be cheaper if it can spend fewer tokens. This won't be the case for all tasks. Opus and Sonnet will also consume the same number of tokens to load context, so the relative price comparison depends a lot on where the task is context-heavy or reasoning-heavy.
I have some business automation process for some customers that I maintain, where Terra with low reasoning ended up with the best price/performance & latency. With an equivalent Luna with high reasoning, price was usually better, but quality had a higher variance (very much depends on where the thinking takes the model) and latency for high reasoning was worse.
If you use API in agent or your own integration with software, such models make a lot of sense. We use Terra 5.6 for "question to SQL" translation for example. It's good enough and affordable enough. Luna will probably not handle complex queries well while Sol is too expensive.
I've been refused to do cybersecurity (read-only investigative access against the recent brotli cve) with Opus 5.5 when I had full SSH access, so I guess you've got to have a fallback?
i use sonnet to generate answers to help-desk tickets, which it does for significantly cheaper and insignificantly worse than opus.
if you're just talking about use in codex or claude code, then sure, leave it on opus all the time. but for running routine api requests, sonnet is great.
I cannot figure it out either.
I use Astra Ultra to plan and 6.1Sol High or X-High to build and even on huge projects with 10s of 1000s of pages it is fast and competent
Cheap and affordable. Dispatching proper tasks to these models can save you a huge % of your quota. Especially if you _know_ the task is simple (ie add something here or there, or archive a bunch of files and reformat kinda thing)
Maybe it's my tasks, but in the cases you're describing, I'm usually better off going down to Luna. I specifically don't see the gap where "It's too complex for Luna-high" and "It's not complex enough for Sol-low" exists.
I've personally found that most models completely collapse at sub-medium effort. I generally run most of my stuff at medium/xhigh with heavy core tasks @ max
Similar 75% price cut trajectory as GPT-5.6 Sol -> GPT-6.1 Sol. Overall, Terra is still half the price of Sol.
But like the original release of Terra, it's still a weird middle child between the small-model-that-got-substantially-better (Luna) and the large-model-that-got-significiantly-cheaper (Sol). The TerminalBench 4.0 cost-per-task results are a bit confusing and I'm not sure why you would pick Terra over Sol in that case. Overall this release seems like a flex on Sonnet 5.5: a model which is having its own middle child problems between the godly Opus 5.5 and the surprisingly capable Haiku 5.5.
It is having improvements in other fields than coding while being more cost efficient.
If all I want is coding at a cheaper price, then deepseek is a good model for that.
I am more than okay to have a bifurcation of models good for coding and models good for everything other than coding. I strangely imagined it many years ago.
Everyone is suggesting OpenAI to release a model close to opus 5.5 yet it's releasing this. On one hand we can imagine that maybe their training cycle wasn't able to produce this whereas gpt 6.1 terra was made.
But on the other hand, I also now wish to see an OpenAI model which is good for just pure coding, like an absolute beast at it while being cheap (so the opposite of terra)
ironically, one of the few appreciations that I might give to OpenAI if they pushed this model to prod and are facing some pushback as developers like us want/only see coding models. They could've benchmaxed it and it would've been slightly irrelevant. I'd much rather see reality.
At the very least, this new model won't make things more expensive for anybody and would only make use-cases cheaper depending on what they are, so it's a fine deal for me.
I also wonder maybe this is why open source labs aren't targeting this space and why the only focus of AI labs is coding nowadays. Everything else just doesn't seem so worth it from investor, even end user, training perspective.
Maybe this is the reason why we have gotten almost an exponential growth in coding capabilities as that became the most (if not only) capability that truly really matters.
To not wish for discussion to go to a tangent, nonetheless this could also explain why our work as a software engineer is going through a slight crisis of meaning as these AI models have gotten impressive at coding.
As an (aspiring) developer (currently in 1st year college), I don't wish for AI to replace coding even more for me or to push it even more or to suck the joy out of it even further. So I am more than happy for these models to stay as they are and rather make advancements elseplace. Ironic because it is us developers who use AI as well and we want cheaper workflows and so we would also want models to be cheaper and better at coding. There are some fine lines of contradictions and irony in all of it. I'd much rather also have open-weight models of such terra-like models but I imagine that because of the reasons that I mentioned above, we are most likely gonna see it the most focused on coding but still open-weights is pretty close.
The more choices the better (though choice-fatigue is a real thing), I hope I was able to explain some of the nuance I feel about this particular launch.
They're doing its because they recently cut plan usage by about 50%, leaving Anthropic providing about 5x the value on their plans. [0]
The "cost per task" metric is just a weaselly way to sound like they didn't just double the cost for most users, and aren't having their proverbial lunch money stolen. They can no longer claim to provide more tokens, more capable models, or more usage in general. So they made up a new metric where "task" could be defined to mean whatever they need it to mean.
They had to become VERY specific in their wording to present any kind of value proposition that will entice people.
> 5x is calculated in terms of API costs, nobody pays those extreme Anthropic prices outside enterprise users.
Ignore value then. Anthropic still provides about 3x as many raw tokens. The 20x\200 plans provide about 10B for openai versus almost 30B for Anthropic.
Do you really think that Opus uses 3x as many tokens as Sol per task? I don't.
> Most users aren't on their $200 plan.
Fair enough. I should have said most heavy users.
> Most users (including on the $200 plan) continue being subsidized.
They continue to be fifty percent less subsidized, I guess.
Look, I don't fault them for trying to become profitable. I fault them for the weird weaselly business speak about cost per task. It's BS game playing and we all know it.
I'm honestly rooting for them, and hope it all works out. For now though I'm pretty salty that my bill jumped by 150% in one month. Gaslighting about made up metrics just makes me more angry.
People in HN live in a bubble. Plenty of real world use cannot use subscription plans (not even talking about enterprise agentic use) and is absolutely at the mercy of the cost per task metric.
We agonize over cost per task to the extreme for services/functionality built atop of APIs and can absolutely guarantee you that it is the real world metric that counts most right after “is the model good enough for the task at hand,” and furthermore that, yes, in fact, it is quite often that we see certain models/providers easily using 3x or more tokens on average for the same tasks (or even the exact same task) than a supposedly similarly matched model from another provider. Once you become intimately familiar with the thinking traces from specific models, you can preempt this to some extent with manipulations to the system prompt (essentially heading off entire thinking branches) but the nuking of reasoning traces in the name of preventing against distillation attacks has made this harder.
Cost per task, time per task, quality (good luck measuring objectively between models) are the metric that matter. I don't care how much a token costs, I'm not buying tokens, I'm buying solutions.
Once you are inserting models into non-creative (and software development is creative work) it is imperative to start creating your own small evals. Not just to check how different models perform in that equation, but to catch model drift over time.
OpenAI repeatedly keeps hurting user workflows by altering the model generations family for each new version in unpredictable ways. As such, it is not a sane ecosystem to develop in.
If you are confused, that is how this company was always been, remember GPT 4, then they did some weird versioning.
I personally like Terra, it kind of does what I would ask Sol or Astra, but dont have to count tokens as much.
And Luna has been amazing tbh, it takes designs from prototype and in record time is implementing it, so that was a huge win.
But yeah, they always had confusing way of naming products and if you are not following news, you will be confused about what it is.
Looks great on the first chart, but terrible on the Terminal Bench 4.0 where it needs xhigh to come close to 6.1 Sol, and costs more for a slightly worse result.
IMO it's useless because we don't even know what a pelican on a bike is supposed to look like, It's very impressive that that models can make these directly using SVG, but we are way past the "wow" stage and test that can't be gamed and are easy to prove how good the model is in that specific arena are hard to find.
The only thing that pelican have is "models where not trained on this" maybe that true months (or years ago) now it's on every HN post and these AI labs are monitoring HN
Not necessarily, they may have selected benchmarks where Terra performs relatively well compared to those models. Pretty much every model release these days just cherry-picks & highlights the benchmarks where their latest model performs best. We'll have to wait for the complete benchmark suite to have a better picture.
the only things we've got to actually make a judgement on are vibes and token anxiety. vibes generally indicate that more expensive == better, but more expensive == more token anxiety. so you end up using the most expensive model x effort that keeps your anxiety in check. yeah, maybe terra is great at technical documents. i'm not going to use it because i'd need to spend a couple of days writing technical documents to work out if it was better than sol and then what if it isn't?
is opus 5.5 better than sol 6.1? hmm, i really can't tell. maybe claude is a bit more diligent and identifies more issues up front. but sol has this can do attitude and the codex harness just gets on and does it's job done.
Why do Anthropic and OpenAI think it necessary to split their model into 4 versions (Opus, Sonnet, Fable, Haiku, Astra, Sol, Luna, Terra) vs. Google's 2 (Argon, Flash)?
Just one thing about the advert, how can I be the only person here thinking the ad is weird - like for the marketing dep, its a bit weird the ad features the same tram that broke and tragically like ended some people - happened for like the first time ever in 100 years, like a year ago ( 14 months ago) just a bit weird.
Im honestly not trying to be negative, but its kinda weird I kinda feel I should mention it because noone else has?
Im sure literally anyone from southern europe must have thought this the second they saw the ad. It was big news last year.
Am I mad? Just a heads up for PR to always check things like this first when releasing a video? Maybe its just me? Just trying to help
e.g.: "You must not use OpenAI models or their outputs to generate, prepare, select, or supply training material or signals for models that are intended as alternatives to or substitutes for OpenAI models or services".
The results for factual error rate on "difficult" prompts is staggering to me. Particularly the difference between 5.6-Terra and 6.1-Terra. These are wildly different Pareto fronts. Our custom enterprise agent is currently running on 5.6-Terra and it's already very good. I look forward to updating the configuration.
These "Pareto frontier" performance vs price graphs have become very popular and I find them mostly uninformative. Pricing is subjective. You chose the position on the X axis, and which measures to use on the Y Axis.
The error bars on eval graphs are commonly like ±5%. It's fairly common to get out-of-order points when the gap between them is only a few percentage points. If we ran it again, or ran it with minor alterations, I'd expect the ordering to be expected. In general, we don't massage any of our eval numbers and for good statistical hygiene we'll report the first run we do, even if later runs show slightly smoother curves.
Subscriptions provide thousands of dollars' worth of tokens if measured at API prices. While the value was once absurd, it is now slightly lower but still much cheaper.
not gonna lie they did really bad and this drained my usage SO FAST. that model used lots of unnecessary skills and read through way more memory files needed.
They say it has "A significant step up on factuality and coding".
My browser is underlining "factuality" in red. AI hallucinations or hallucinatiality are still a big problem in my experience.
AA has limited 1017 benchmarks so some terras versus 1017 Max:
max - 3 points smarter, 50% more expensive
xhigh - same score, 25% cheaper
high - 1 point dumber, 50% cheaper
Yes, V4.1 Flash 1017 is my daily driver cloud model. Copes fine with what I'd consider difficult brownfield development, like fixing bugs in vLLM prompt cache and coming up with small/sane overlay patches for them. Maybe it's disappointing for people who want to one-shot a game or website from a single prompt; I haven't tried that.
It's nice that their peak-rate hours end quite early in the day UK time.
You can do real world development with weaker models than Deepseek Flash.
I still use SOTA for the main model though, but also I'm perfectly aware it's mostly out of FOMO.
Deepseek is more than enough for default subagents though. Any well defined and verifiable task and there's absolutely zero difference between SOTA and good-enough models. Used to be just speed, but Deepseek is fast.
That terminal bench 4.0 result is... bad? Terra XHigh more expensive than Sol Low and still does worse.
It's very likely this is a next checkpoint of GPT-6 "Sol", over-optimised for the web chat and offered to Codex subscribers as a "please go away" offering.
If this was half the price, I'd be very interested, but at 1/5/0.05? Feels like another panic release because they can't get Astra 6.1 out of the door...
People keep asking OpenAI for a mode that's as good as Opus 5.5, and they release this instead. Interesting strategy. Let's see how it works out for them.
They don't have anything competitive with Opus 5.5 yet. There is a supposed model in the works, "Bel" last time I paid attention to gossip, that is thought to be better, later this year. Allegedly Anthropic caught OpenAI off guard and / or their different training approaches ended up working better than anticipated.
My expectation is Fable 5.5 is next overall improvement, and my bet is OpenAI's "Bel" will be the next competitive one (or, their lack of safety culture will bite them/us in the process).
There’s also a chance they surrender the dev market altogether to Ant and just focus on the “normie” segment with dots, computer use, excel etc. In which case they won’t have anything competitive to opus-5.5 anytime soon for SWE.
This release is interesting for enterprise at API rates.
At least unless the Terminal Bench 4.0 score is more indicative than the other ones, because that one shows it clearly worse than 6.1 Sol for both performance and cost.
Anthropic models are comically overpriced. They also gobble up an obscene # of tokens to get the same result as others. The price delta is literally 5x.
Taking the benchmarks (as reported by OpenAI) at face value, why wouldn't they show Opus 5.5 on the OS world benchmark, and show no Anthropic models at all on the "factual error rate" benchmark
Right?! I think sol has some obsessive tendencies, i literally caught it the other day spawning a sub sub agent doing a final fourth test for a piece of code that isn’t even critical say in embedded or similar, now I have to instruct it to only test once!
Note that while that page is from Pangram, the studies on it are not. If they are excluding studies that show Pangram performs poorly in realistic conditions, please share those studies!
This is incorrect. You cannot identify the presence of text watermarks if you do not have access to the models so that you can spot where expected tokens diverge from actual tokens based off the weights. That does not preclude Pangram from being able to reliably identify LLM-generated text from models whose outputs they’ve trained their own detection model upon.
I'm very confused about their model market position/naming schema.
First we had:
> Luna (Small), Terra (Medium), Sol (Big).
Which was fairly smooth after you got used to their planets.
But then they added Astra as the "Bigger," which was mostly still self-coherent:
> Luna (Small), Terra (Medium), Sol (Big), Astra (Bigger)
But then things go confusing, they released Sol-6.1 which sometimes is superior to Astra, and discontinued Terra which made it look like this:
> Luna (Small), Sol-6.1 == Astra
But with Sol-6.1 costing the old Terra-6.0 price, while, as I said performing as good as the old "Bigger" model in many non-3D/UI situations.
Now they're re-introduced Terra, as a 6.1 release, making this even more confusing.
Did Microsoft's chief branding officer move over to OpenAI?
I work at OpenAI and I can try to explain our logic.
In general, LLM performance is incredibly multidimensional. There's no guarantee of a fixed ordering where the model that is best at math is also the model that's best at jokes. Despite that complexity, much of the variation in performance is explained by just a single variable: model size. In general, models that are bigger will (a) be more performant / reliable, (b) cost more per token, and (c) have better recall of esoteric facts. As you note, our models are ordered:
Luna < Terra < Sol < Astra
Even today, I'd recommend this as a general starting point for picking models. If you care about quality only, pick Astra. If you're a bit price sensitive, pick Sol. If you're very price sensitive, pick Terra or Luna.
Of course, model size doesn't capture all variation. Models that are trained later might have new data or new techniques or new architectures that earlier models lack. You can see this with models from all providers - Sonnet 5.5 outperforms Fable 5.1 in ways, and Gemini 3.8 Flash outperforms 3.1 Pro in ways.
In the case of GPT-6.1 Terra, I'd say Terra is probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things. For API developers, it's a nice midrange option if you want something cheaper than Sol but a lot smarter than Luna (especially for vision / computer use / communication). For subscribers, it's not a bombshell release - if you're hitting usage limits, maybe it helps you go a bit further, and if you want clearer communication or faster computer use, it can be a nice option to select.
If you have suggestions on how to name our products better, I'm all ears. Our intention is to make things clear and simple, even if we miss the mark sometimes.
Unfortunately, a lot of the complexity is somewhat irreducible and due to to the actual complexity of the products. Scaling model size is different from scaling reasoning effort which is different from scaling multi-agent count. So if we just swept all of those under a single name of "more IQ", it would make the naming simpler but the experience much worse, as important details would be obfuscated from end users.
Do you then recommend Sol for technical documentation?
What's confusing to me in addition to the naming is that there are no published guidelines I can use, as far as I can see. Put yourself in the user's shoes. On what basis should one pick a model for a given task?
I can understand smaller models being dumber and faster. That aligns well with scenarios where you are willing to compromise; e.g. you're classifying content or asking a model to format text into a table.
But when you want "maximum intelligence, cost and time be damned", there's the implication that bigger models are smarter on a linear scale, and every provider's model release naming is currently conflicting with that expectation.
This isn't just a naming problem, but a communication problem. If I knew that Astra was the best at coding but a bit bad at writing, and Sol was almost as smart but better at technical writing, then that's the kind of thing I expect to be documented.
One possible way: A simple matrix with horizontal columns like "writing", "coding", "math", rows are the models, and each box is a number of stars from 1-5.
Instead, this sort of thing becomes lore, shared by word of mouth.
> Do you then recommend Sol for technical documentation?
I don't know, to be honest. Technical documentation is such a multidimensional skill, and I also don't know your budget. I'd imagine Terra might have the most readable prose, but Astra might be more likely to catch bugs or mistakes to highlight. If the documentation is about a more esoteric language or situation, Astra might have a greater edge than if it's a common language or situation. If you have a long list of documentation guidelines, that might also give more of an edge to Astra. I wouldn't want to make a recommendation without doing a lot of testing, and I haven't done a lot of testing.
Some food for thought, if someone from oAI can't make a direct recommendation for something as direct as "Technical documentation is such a multidimensional skill" - how could we know assess what model to use?
To elaborate, I can certainly ask each model to write a doc about something I know well and tell you which output I like better.
But to assess whether my preference on my problem is going to generalize across other people, other programming languages, other documentation types/formats, etc., and then assess statistical significance is a much broader question that I personally don't have the data to answer. I'd rather make no recommendation than a poorly informed recommendation.
I use Astra 6 Extra High for everything.
It’s overkill most of the time, but frankly all these models are “cheap”, and I’d rather waste compute than brain cycles trying to discern which dials are best for each individual as-hoc query.
> Put yourself in the user's shoes. On what basis should one pick a model for a given task?
Create a realistic eval suite for your use case using your actual data that you manually label, with easier, more complex, rarer samples, and run each model you are considering against it and decide based on the usual confusion matrix, cost, latency, or whatever metrics matter to you. There's really no other way to know – and you might be surprised that cheaper models can often perform your task(s) relatively well at a fraction of the cost.
> Create a realistic eval suite for your use case using your actual data that you manually label, with easier, more complex, rarer samples, and run each model you are considering against it and decide based on the usual confusion matrix, cost, latency, or whatever metrics matter to you.
I’d imagine there’s only a tiny percentage of users that can or want to do this.
> Do you then recommend Sol for technical documentation?
Every model I've tried is just abysmal at this, for what it's worth.
If any do it well please let me know and I'll give it a spin.
My thoughts exactly.
I recently experimented with some image generation models and I was surprised at how much context/prompting is needed to get something that doesn't look Lovecraftian. There was seemingly no downside to adding more and more detail up to the limit. Prose sits between code and images in terms of fuzziness and I wonder if promptmaxxing is the winning strategy for technical documentation.
I've been thinking that fine-tuning a model to write differently shouldn't be that hard (tedious, perhaps) and if I had the resources I would consider doing it myself.
You can start with Artificial Analysis benchmarks but, then you need to test in the real tasks anyway.
I assume having a person write words for other people to read is not feasible?
Humans cannot be in the loop for autonomous work. I have agents review code and design/architecture documents, write alert messages based on rules, write analyses (e.g. outage incidents), and many other things. "Having a person write words" would mean not using agents, defeating the whole point.
I do not use agents to write on behalf of myself. I write my won technical write-ups and my own slide presentations. A lot of people don't; and it shows. There's much that agents can do better than humans now, but not everything.
This isn't technical documentation though, it's just notes the agents are keeping for each other. I think that is a very different task
It's human-facing.
> I'm all ears. Our intention to make things clear and simple, even if we miss the mark sometimes.
Thanks. I think the very response you wrote here is a better form of marketing for the software engineers than what openAI puts out there. When everything is groundbreaking and everything tops charts, nothing is.
I suspect a lot of target demographic (developers) prefer simpler and honest communication about not just the models high points but low points too. With the linked Terra post, if I saw 1 chart that clearly highlighted the usecases for which Terra is _worse_ than higher tier models, I would be more convinced overall.
Everything trying to occupy the top spot in my brain's limited "top model" representation space is not helpful.
While I generally agree that it'd be nice to just have "here is the Pareto optimal model for your use case. Here is the specifics we optimized each model for" so users don't have to do benchmarking and research as much, they're getting much better at explaining. It's basically what the commenter above said.
The website linked starts out with:
> We’re introducing GPT-6.1 Terra, a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing. A significant step up from GPT‑5.6 Terra, it excels at the communication and computer-based tasks that fill the workday, from drafting proposals and interpreting technical drawings to filling out forms and editing videos.
If you go father down, it has graphs and details with specifics of this, including:
> We measured GPT-6.1 Terra’s computer-use capabilities... At medium reasoning effort, GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra at about one-third the cost. It even performs on par with GPT-6 Astra at only one-tenth its price per token.
and
> On MMMU-Pro... It even nearly matches the scores of our most capable frontier model, GPT-6 Astra, and GPT-6.1 Sol at much lower cost.
The 4-tier naming is honestly as clear as you can get. I have no idea what people are so confused about.
Terra and Luna are dirt cheap and good enough for 50% (or more) of the typical junior dev's workload, and great for routine work. Sol and Astra have so much written about them at this point that it should be obvious what those are good at. The large/gigantic models for things that need full intelligence and slightly more branching thinking.
The obnoxious/confusing thing is that models are released individually rather than as a family.
Luna < Terra < Sol < Astra, fine. But does Terra 6.1 beat Sol 6? Does Fable 5.1 beat Opus 5.5? If I build something around this Terra 6.1 release, then when Luna 6.2 is released in a month, should I swap?
I understand the confusion, but isn't this kind of similar to asking does the iPhone 14 Max beat the iPhone 16 Mini? At least with phones there are hard specs we can use to compare, but in a lot of cases the answer will be "it depends"
At least there is a simple system - the words 'Mini' and 'Max' are intended to refer to the size of the item (that one is miniature, and the other is their maximum size).
If they were called iPhone 14 Luna and iPhone 16 Sol then this is less clear.
Or on the flipside, if the models were called GPT 6.1 Mini and GPT 6.1 Max, nobody would be arguing about the names being hard to grok. The words 'mini' or 'small' are more descriptive than 'Luna' or 'Terra'.
The problem with that is then you would have GPT 6.1 Max with max thinking which is different from GPT 6.1 Small with max thinking.
If you look back the naming problem started around the point that OpenAI started doing reasoning models (o1, 4o) where they had to describe this fork in model behavior that let a smaller model think longer and be smarter.
You can come up with naming schemes that don't clash, but there's something to be said for this approach where the thinking level is like you describe and the model name requires you to learn that Astra is the largest model.
> does the iPhone 14 Max beat the iPhone 16 Mini?
Sure, in a hypothetical world where the latest Max phone is two generations older than the latest Mini phone. In which case now you have to evaluate not only screen size but ALSO specs and component age and software support etc. How badly do you need the bigger screen? Etc
If you are comparing the iPhone 16 Max to the iPhone 16 Mini, everything is a lot easier.
> iPhone 16 Mini
I wish
Yeah, but shouldn't then all 4 tiers be released at once? So we get 6.1 for all Luna/Terra/Sol/Astra?
No. It’s not similar at all.
It's more like if Apple called the iPhone 16 the "iPhone 14 16 edition"
I’d rock an iPhone 1942 Edition.
Don’t forget to refill the kerosene tank!
Run your own evals then if you have this much of a need to "correctly" choose. That'll make it clear.
I'll just stick to mostly Sol on low/medium/high and start experimenting with Terra tomorrow onward to see what it does noticeably better than Luna/Deepseek flash.
Any pareto frontier model, especially from the leading labs is guarnateed to be good enough for whatever your expectations are at this point. A quick few day test should be enough for anyone to make their own conclusions.
Right, having products so poorly described that a customer needs to run their own evals to choose well is exactly what is being complained about.
Or I'm going to give up and not use any. Consumers aren't perfectly rational. This isn't like toothpaste, where, despite being overwhelmed at the store, I still need to brush my teeth so ahhhhhh fuckit and just pick a toothpaste out of the 30 available. The alternative is I keep using Google search and think AI is stupid. Or just use Facebook's thing.
The frontier toothpaste is the one with bioglass 45S5. Especially if you have sensitive teeth or your teeth are prone to caries.
> > But does Terra 6.1 beat Sol 6?
> Run your own evals then if you have this much of a need to "correctly" choose.
No. OpenAI want my money, it's on them to communicate properly.
What would you want to see instead? Not sure about you but my needs are specific enough that I don't think OAI has evals for them, and I wouldn't want to give them that data even if they wanted to run it for me.
“ChatGPT 6.1 Max” would be simple.
Then you're going to have ChatGPT 6.1 Max Low and ChatGPT 6.1 Max Max.
How does one more-properly articulate "it depends"?
On what?
> The 4-tier naming is honestly as clear as you can get. I have no idea what people are so confused about.
I wouldn't necessarily say that. Atmospheric mass may be obvious to us, but I suspect it's not nearly as universal knowledge as one may assume.
And I actually assumed at first that they were based on distance from us (terra < luna < sol), so I kept confusing terra and luna.
It’s totally clear. About as clear as mud.
I think the main problem is surfacing two topics:
- Understanding how effort and class interact (Is a high effort terra above a low effort Sol? Are larger models mostly always better regardless of effort?)
- understanding when (if ever) a new model tends to be better than an older but higher class one. Is an old astra better than a new terra?
If increasing effort and increasing model class produce improved results in a different way, I miss making it explicit in which way. In other words: What kind of issue warrants going up in effort? what kind of issue warrants going up in model class? What kind of issue is unlikely to be solved by either?
OpenAI itself cannot answer these questions. A prospective price-sensitive (or speed-sensitive) user has to experiment to find the cheapest (or fastest) model that is sufficient for each task.
Right. Because they don’t know. They’re just pumping out crap desperately trying to stop hemorrhaging billions.
>effort / performance curves by model
Oh, easy:
https://deepswe.datacurve.ai/
Artificial Analysis has their own curves which are pretty similar.
That's good, but they haven't updated it for a month. It is missing all of the current frontier models.
Seems to be missing all of the 6.0 and 6.1 models?
> Is a high effort terra above a low effort Sol
You can literally see this in the charts from the article
Since internally you probably have a fairly good intuition for what these models do well, what about assembling a benchmark suite (with multiple dimensions) that produces results that show that intuition, but in a 'somewhat objective' way?
That could solve the 'more IQ' problem, because you'd actually have different dimensions (e.g. for 'raw intelligence', technical writing, emotional understanding, etc) that you could use to communicate the differences in models better.
(Of course people would instantly start complaining that the benchmarks numbers are probably fake etc, but it is what it is.)
I think different domain benchmarks already exist and are widely used.
True, but afaik none that an AI company itself uses to say 'this is what we think showcases the individual strengths of our models the best'.
Internal evals are still running but GPT-6.1 Terra is an AWESOME release for anyone using the API in agentic production applications where cost matters.
Personally, I pretty much exclusively use claude for coding work (sorry!). Our application is running almost 100% on OpenAI API (except for some fallback models). It is the best API on the market. The models are excellent particularly for any application that is cost sensitive.
It's not gonna happen but:
> Despite that complexity, much of the variation in performance is explained by just a single variable: model size.
please release model size :p
What makes this even more confusing is reasoning effort.
You have Luna, Terra, Sol and Astra, each with 4+ reasoning levels. That's ~20 options to choose from.
On a repetitive task like customer support or something, where you have decent benchmarks, you can plot cost versus performance, but you can't really do that on coding.
This was lovely. Can we have more of this style of communication?
Instead of benchkmarks it'd be useful to get a sort of Radar chart across measured capabilities like computer use, communication, coding, etc.
We're probably headed to further specialized agents where a single ordering or even model size won't mean a lot.
I actually do have a recommendation, which is to do something like a new mode (i know i know): Auto xlow Auto low Auto med Auto high Auto xhigh Auto max
With an optional flash mode.
Then as you release models, you pick the model and thinking level that you believe is pareto optimal for cost/performance.
So this auto updates each release, like Terra was probably off this list entirely, but maybe this would appear now in Auto medium or something like that.
I know from looking at charts that each model seems to have sweet spots (like I always use this model in this thinking level and if i want to go cheaper i go here). If you picked those for us, it would be interesting. I do think you should keep the model families for those wanting more consistent experiences, but for many devs they may want to just pick auto something and have it update over time. I think nice if it actually just said in parenthesis what it was, something like: Auto Medium (Terra 6.1 High)
I am not sure if everyone would agree here though. I'm obivously advocating for more naming schemes, and i understand the irony.
Just to be clear, I think the current mapping for OpenAI is something like:
xlow = Luna Low
low = Luna xhigh
medium = Terra medium
high = Sol medium
xhigh = Sol xhigh
Max = Astra something?
"In the case of GPT-6.1 Terra, I'd say Terra is probably better than Astra and Sol at technical communication" -- you're doing the thing here too, underspecifying your own models: 6.1 Terra is better than…which Sol, 6.1, 6, 5.6, just the latter two, all of them? This 2D bluriness is a big part of the confusion.
it depends on the task and it's hard to say
Despite Opus 5.5 being newer and "better" if you read benchmarks, I've found Fable to be better for some things (subjective stuff)
I like to use OpenAI models for ingesting a customer brief, usually a paragraph or two and then a read of a several customer files (eg spreadsheets and docs etc) and turning it into a system design blueprint. I'm currently using Astra for this phase but it's incredibly expensive relative to the other tasks that come after this. Would you recommend I stick to Astra or can I downgrade to say Sol or Terra for this job?
I'm just worried the system design won't be as good if I choose a lesser model so to speak.
Edit: typos/grammar
Please can I ask about one parameter you didn't mention here? All of your models have both a reasoning effort setting and also a non-pro vs pro setting. I have some intuitive feel (from experimenting) for the difference a larger vs smaller model makes, and why that's different from adjusting reasoning levels, but I have no real feel for how pro vs non-pro is different in behaviour from adjusting effort.
I think you can think of Pro approximately as running multiple instances of a non-pro model, than giving all their outputs to a final judge model which combines them into a better one.
I use OpenAI models all the time (via Codex) and I've never seen this pro toggle. Where are you seeing that?
Appreciate that these things have orthogonal measurements; with that in mind, given that the name signifies size, what does version number signify?
I.e. are all version numbers independent and Model X 7.5 is not necessarily of the same generation or cut off as Model Y 7.5? Or do all of your 6.1 models have something in common - cut off data, generation, baseline technology, approach, architecture, or something else?
Thanks! :)
These days, version number signifies a few things at once: the order of release, the relationship to other releases, and the magnitude of improvement. So something like 6.1 would indicate it was launched after 5.6, may share some similarities to other 6.1 models in terms of training data / personality (which is why it's not 5.7), and isn't a giant enough leap to merit the headlines and expectations that would come with a 7.
Back in the day we used numbers to also sometimes indicate size, which could make it extra confusing: e.g., we launched GPT-4.1 (an incremental update on GPT-4) after GPT-4.5 (a new model that was a big jump over GPT-4).
This makes a lot of sense.
It’s probably just unavoidable. Like when Apple release a new MacBook Air first and it is (in some ways) faster than the old MacBook Pro.
I think the planets thing is a bit confusing though. I hadn’t really thought about it being based on the size of the planet. I assumed it was based on how far away they were with terra therefore being the base one.
This kind of heuristic is actually extremely useful - and mostly what I try to develop over a couple weeks as the new models come out. I would love if we could get a fairly standardized take like this on the regular.
Of course we have far too many benchmarks (of questionable utility), but it would be helpful to get OpenAI's take on where it slots in, and how the landscape changes as these get rolled out.
It would be interesting to know which model is internally viewed as being best at jokes, I have a personal humor benchmark and so far models will either steal jokes (from comedians like Maria Bamford), or write something very unfunny.
When I was first learning to have agents interact I had a little script to orchestrate a model telling a joke, a critic collaborator helping develop material and an audience emoji rank them.
A favorite…
Prompt: “Tell a joke about mother in laws”
Gemma2 response: “What’s the best kind of mother in law? A dead one.” 0_o
Straight up - why not just name them Cheapest, Cheaper, Expensive and More Expensive? Cost seems to be the main driver of the intuitive difference between models.
Same reason that Apple doesn't name its series of phones the iPhone cheap, iPhone regular, iPhone semi-expensive, and iPhone expensive. It's poor branding, isn't easily extended to future models, and gets confusing when comparing across model generations, where an iPhone 14 expensive might be cheaper than an iPhone 16 cheap.
I love the current naming scheme and I am excited to see the astronomy trend expand in the future. I hope to see asteroids, binary stars, and solar systems at some point soon, as well as neutron stars, black holes, and eventually galaxies, and clusters of galaxies.
> Models that are trained later might have new data or new techniques or new architectures that earlier models lack.
This seems like one of those cases, where CalVer would be appropriate, and you likely have ideas what you aim for while training, so technical model names (ids?) could include aliases to most things you said. Like 2026q4-xlarge-researcher, 2026q4-medium-humanitarian, 2026q4-medium-coder, 2027q1-small-analyst etc.
> important details would be obfuscated from end users.
I don't expect the vast majority of your users care. I would guess they want a rough proxy for "cost" and will assume cost=>capability are roughly linked in a linear fashion.
it is very possible that the users weighted by spend do, and I'm therefore simply wrong. But, 2c
In your own response you mentioned GPT-6.1 Terra, but I've seen GPT Terra-6.1. I would start by the part that is the most powerful:
GPT 6.1 X > GPT 6.0 X GPT Y Astra > GPT Y Sol > GPT Y Terra > GPT Y Luna
This order is very important.
The length of your explanation shows just how bad of a naming scheme this is.
What prevents OAI (and Ant for that matter) from releasing new version for all models at once?
Instead, we got Opus 5.5, Sonnet 5.5 and Haiku 5.5 released weeks apart, and so are Sol 6.1 and Terra 6.1?
My guess is nothing, and they're currently just releasing them individually as soon they're ready instead of all at once so that they hit the market more quickly. AI moves quickly enough that a delay of a couple weeks could result in significant lost revenue.
Lower models typically use distilled outputs from higher model to improve .
all training runs have to at some level share the compute capacity ; while they have more than enough to run many large runs concurrently, they don’t have enough compute to run everything concurrently.
> I work at OpenAI and I can try to explain our logic [...] general, models that are bigger will be more performant.
> If you have suggestions on how to name our products better, I'm all ears. Our intention is to make things clear and simple, even if we miss the mark sometimes.
How about naming the models:
"Small", "Medium", "Large" and "Very Large"
Or refer to them by parameter count like other companies (e.g. this is 6.1 5T)
If I have a clothing store, and rather than name the clothes sizes Small, Medium and Large I decide to call them Urgon, Planfar and Fangor I shouldn't be surprised if people think my size naming is confusing or silly. And IMO it's not a great explanation to say "oh it's easy, if you are skinny then you should get Urgon, but then if you are a bit more of a medium build you should get a Planfar, but if you want something a bit baggier than Fangor will really work out for you"
The problem with small, medium, large naming is that they don't last. In a few years our names would have to advance to labels like giant, behemoth, gargantua. This is different than clothing, where we know that in five years humans will be most the same size as today.
As a point of comparison, GPT-3 was launched in a series of small, medium, large, and XL. Back then, XL meant a mere 1.3 billion parameters!
> In a few years our names would have to advance to labels like giant, behemoth, gargantua
You wouldn't - you would just put the version before it - 6.1 Small, 6.0 Medium etc. Not that complex.
Think of an iPhone 18 Pro - No need for 'iPhone 18 Even Better Super Pro' because the 'Pro' is relative with the version/generation.
Reminds me of astronaut sizing charts for ... their waste management devices.-
Use colloquial terms: Big, Fast, Genius, Affordable, etc.
> If you have suggestions on how to name our products better, I'm all ears
Terra 2026.2
Would you like to recommend a model that does not produce errors and can be relied on like a traditional computer, please.
You certainly know that it happened not by accident that traditional computers are producing reliable results, a lot of work went into this.
Computers must produce reliable results, they are simply not useful if not.
People will understand that at some point, but a lot of damage already has been done.
What a dumb comment. Go on your anti-LLM tirade elsewhere.
Is there someone I can email directly about fine-tuning - all my models have been queued for 24/48 hours and your customer support seems non-existent
Fwiw, if you have to explain your product's name to the customers who just want to use it, that's a sign the products are confusing.
Tiny/2026-11. Medium/2026-6. Big/2026-10. XL/2026-01. It may not be sexy, but people know exactly what it is and what to use.
> Our intention is to make things clear and simple, even if we miss the mark sometimes.
Well you definitely missed that mark extremely badly here.
What would you have like to seen instead?
This is kind of like Apple with the iPhone. You need 3 model products max. 3 is easily understood and people can be bothered (if they care) to track the difference. More than 3 is a mix of being redolent of ADHD (viz bad PR) while also being annoying for consumers to track the difference.
I can’t tell you the difference between the 10 different versions of the current iPhone and I can’t you the difference between the 5-6 different models put out by the Anthropics/OpenAIs of the world. At this point I just try to track the flagship and then the cheapest one or worst case I just use the flagship.
I’m sure internally there’s data showing people want choice and options but decision fatigue is of course very real.
Very fair. There's always a tradeoff in the number of products, and it's hard to make everyone happy. My suspicion is that the optimal number is probably a bit higher for developers on our API (who often build their own internal evals) than for individual users (who don't want to waste time re-optimizing every month). We're excited by Terra's uniquely fast computer use, but if it traction is poor, maybe we'll go back to 3. It all depends on the market signal - we only want to ship models people find uniquely useful.
If there's no fixed ordering, can we have model names that describe the capability?
Have a model alias called "math", another "jokes". Quality labels/date suffixes to these obvious labels. Repoint as needed.
Something for people who don't want to play the cute "which part of the solar system is trending this week" game.
I promise you this would become a giant mess. No one wants to pick GPT-good-at-writing-but-not-in-documents-and-decent-at-instruction-following-but-has-some-known-quirks-to-be-aware-of-p.s.-good-at-knock-knock-jokes-MAY-2027.
Even "writing" as a single concept is very multidimensional. There's fiction writing, nonfiction for experts, nonfiction for learners, deep research, quick lookup Q&A, copyediting, writing in documents / artifacts, instruction following about style, and default quirks (e.g., em dashes, negation, loving the word "tapestry"). Some models are great writers but less diverse, meaning people prefer the first story they read but disprefer the tenth. How does one label such a model?
> No one wants to pick GPT-good-at-writing-but-not-in-documents
But this is exactly what we have! Except it's worse, because the model usage recommendation isn't in the API, it's in some random person's X advice that's no longer current this week.
I understand AI companies wash their hands of this problem because it's hard, and they themselves barely understand the true capability matrix of the models they're selling.
I just don't think some effort to categorize models and expose that to users would be messier than what we have -- model selection driven by influencer voodoo bandwagons and analogies to literal astrology.
What do you understand by “technical communication”?
naming suggestion: small, medium, large, x-large
it's almost like ya'll let the AI name it ;)
What if GPT 6 small is bigger than GPT 5 large?
Why would that matter? They don't disclose info like parameter sizes to begin with, that's just going off rumors/"leaks".
That’s totally fine since it’s a different family. The new iPhone mini is much larger than the iPhone 4 but its mini in its current family line.
> If you have suggestions on how to name our products better, I'm all ears.
Hire an actual marketing person instead of letting engineers / ML researchers do things they're clearly terrible at ?
> If you have suggestions on how to name our products better, I'm all ears.
For some reason, this is just hilarious to me. Idk if you're being serious or if this is passive aggressive but the thought of OAI actually taking advice from a random HNer and basing their naming scheme off that is peak.
This is one of the last bastions of semi-serious discussion on the internet, and it's full of incredibly high level engineering talent.
Where else would you go? I'm honestly scratching my head thinking of someplace that wouldn't run you out on a rail, but that would give serious feedback. Kernel mailing lists are out, maybe there's something like that which I can't remember off the top of my head, but...
Also, nomenclature is incredibly hard. There are only two hard problems in computer science: cache expiration and naming things.
That’s the issue - this isn’t an engineering problem, it’s a general consumer problem. It’s what focus groups full of regular people are for.
But regular people are not the consumers of OpenAI's API - they consume "Chatgpt" and don't care about model nomenclature.
The primary users of any API are the designers and engineers of the applications that use it. How many laypeople know what Postgresql is? How many of the applications they use run on it?
We aren't talking about the API. We're talking about the model names. If someone fires up the OpenAI website to ask it a question, nobody knows what the names mean. That's what normal people do and are probably super confused at what is going on.
I don't understand what's so confusing about this.
The lineup of models is still the same.
There are four models: luna, terra, sol, and astra.
Pricing increases for each level. This new 6.1 Terra is 50% cheaper than Sol 6.1.
They are lowering prices for newer releases, so yes if you compare Sol 6.1 to 5.6 Terra it won't look right, but why are you doing that? Look at the same release version for comparable numbers.
Each model improves in successive releases but they maintain pricing and capability coherency within a release version.
Yes, Sol 6.1 is supposed to rival Astra for some work, which is what you'd expect from a later versioned release model against a prior release bigger model.
Assuming they release Astra 6.1 at some point, I expect it'll be better than Sol 6.1, at a higher price.
I don't understand what people are asking for here? No reductions in pricing? Different model names on every version release? Not releasing any models for a version number until they're all ready?
Hey! I think it's confusing because there is a mix of three components:
All 3 of these are a one-dimensional balance of price/speed, and capability (Loose terms). The confusion is in how these overlap, how they're announced/marketed etc. So you go from 1 dimensions to 3; you can visualize it as a box. For a given task, how do you choose where in this box you select? Which regions are your favorite? Can you map each region of this 3D space to specific tasks? Label them with intuitive names?Given the operations of the model are opaque, can you see how someone (most people?) might be confused?
But what's the ask? This is the cleanest way I see of them doing it.
I see a ton of criticism and not a single person coming up with anything better, unless it's to take away options, like removing effort level.
What about buying Intel CPUs? A car? Electrical components?
If you're not just using this at a consumer level to ask about cooking recipes (and if that's your use case, you really don't care about API pricing, or any of this, and can just use the default), and are using it professionally, then this seems like a low bar of effort to understand what you're buying.
I do grant you the relationship between model and effort level has an unclear relationship, but there are also tradeoffs in most complex technical products that are not always clear; this is exacerbated by the fact that this product is stochastic.
Why are engineers using this as a professional tool expecting this to be bought like a consumer gadget and then blaming the maker?
“The ask” is give me a clue how to choose the best settings for a specific task. Model releases happen so often that the calculation keeps changing. Maybe even OpenAI doesn’t always know the answer.
This morning I asked GPT which models and settings I should use to inspect my VM dashboard with computer control. It said: - GPT-6 Astra at Medium - GPT-6.1 Sol at Medium (alt) - use Luna only when the route was already known.
But I had already completed the task twice with Luna and it was 5.6! It navigated through a few pages of my dashboard, pulled up the 30-day usage chart, analyzed my CPU and RAM, and explained why I didn’t need an upgrade. I’d give Luna 5.6 an A on this task.
Maybe OpenAI should give us a “Tiptronic” selector: recommend the model and reasoning level based on the prompt, while still letting users override it.
> “The ask” is give me a clue how to choose the best settings for a specific task.
What are you talking about? If it's for some slop you vibe check it. If it's for a serious effort you write an eval. Same as it has been for the last 3 years.
These names are not just for engineers.
Why is it ok for engineers to have confusing names and ways to interact with something?
How am I supposed to decide between these options: https://imgur.com/a/bXHWowL
How am I supposed to keep up with the changes every few months?
You are showing a selection from some third-party interface that mixes together models from six different suppliers: Mistral, Kimi, OpenAI, xAI, Anthropic and Google, with a bunch of old obsolete models.
Are you presenting this as evidence that OpenAI's naming scheme is bad?
If you don't want to spend the time to know the difference between all these different suppliers and their models, then why are you using whatever this program is that gives you so many choices, most of which are bad? Why is this OpenAI's fault?
"How am I supposed to keep up with the changes every few months?"
No worries, you can always ask a AI.
The OpenAI employee above says that some smaller models of the same generation are better at certain tasks than larger models, regardless of cost difference.
They should have a guide saying "Terra is best at technical documentation and communication, Astra is best at large software projects and research, Sol is good for development" or something like that.
Effort (as it's used in the context, how many "thinking" tokens to burn) is still a free variable. It's not always shown to the end user but you can set it independently for any of these models.
After that you just have model "strength" and version number. But of course strength is only comparable at equal version number and the two axes are almost parallel.
Anthropic also does this, perhaps to a lesser extent.
It's confusing because the models are released one at a time rather than as a family.
Let's say I have some task that Luna 6.1 cannot do, but Terra 6.1 can. So I build my product around that. In a month when Luna 6.2 is released, then I have to figure out whether it's worth to evaluate and swap or just wait for Terra 6.2.
I think a lot of confusion comes from the discombobulated rollout of these models as well.
They released gpt-6-sol (alongside astra and luna but no terra) and 7 days later released gpt-6.1-sol making gpt-6-sol basically DOA before any real adoption. Everyone knows it wasn’t just better RL/post-training, it was a completely different model underneath. Now they randomly reintroduce terra.
I’m sorry, but if you’re not even a tiny bit befuddled by this then you must be taking whatever OpenAI is having.
Part of the confusion is that it's not linear capabilities. One model is not strictly "better" than the other. They have orthogonal capabilities.
I think the names are more of a reference to price. Luna is always cheaper than terra which is cheaper than Sol which is cheaper than Astra (for the same model number).
> I think the names are more of a reference to price. Luna is always cheaper than terra which is cheaper than Sol which is cheaper than Astra
Per token, yes. But per task it's not even necessarily the case depending on how much thinking is needed.
That's why "They have orthogonal capabilities" matters. You need to find the one that's best suited to your particular task.
But the model names are linear by price.
It's like Intel i3/i5/i7/i9, or nvidia 50/60/70/80.
It is meant to give a rating inside a generation. Different generations are not necessarily comparable. an i7 is better than an i5, but not if you compare an i7 7xxx with an i5 14xxx.
A nvidia xx80 card is better than a xx60 card, but not if you compare a 1080 to a 5060.
The only difference for OpenAI is that not every generation has every model.
I guess it's the same as Opus 5.5 being in a way better than any Fable.
That's definitely a problem too.
But minor credit to Anthropic, at least they've mostly stuck to their four model names (Haiku, Sonnet, Opus, and Fable). You mostly know where a model is MEANT to be positioned from the name alone. OpenAI cannot keep it consistent from release to release.
But, yes, currently Opus Vs. Fable has the same issue as Sol-6.1 Vs. Astra.
I don’t think opus-5.5 is better than fable-5.1. I think opus is better overall because it’s much faster and cheaper than fable but if speed and cost is not a factor fable it’s probably just as if not slightly better than opus, especially at front-end design (tho the difference is marginal and no one should be really using fable day to day due to this alone).
I don't know why this is so hard to understand.
It's like iPhone 17 being faster than iPhone 16 Pro. Because it's newer and benefits from advances in chip architecture, despite being a lower model tier. Sometimes different iPhone models are released at different times.
Yeah this isn't intuitive at all.
Both companies have screwed up their naming royally. Maybe that's intentional - I'm sure it doesn't hurt Anthropic to have people using Fable a bit longer than they need to and racking up a huge bill in the process.
They're always very clear about it when a new model comes out. You have to click away a banner to continue using the service. There's not much they can do to make it any more obvious without it being a major hassle for all the people who understand versioning.
It is: Fable 5.5 will be better than Opus 5.5. Opus 6 will be better than Fable 5.5, Fable 6 will be better than Opus 6, etc.
You're comparing different generations, any 5.5 models could be better than any 5.0 model.
Really? Haiku 5.5 is better than Fable 5?
If you daily all these models (and have a modest budget, not some >$400 corporate allowance), then it's pretty obvious that Luna is small and fast, and Sol 6.1 is the big one for jobs that need significantly above average intelligence, comparatively to a human engineer. I use both constantly. Astra is really not relevant for the average dev except to do some moonshot task with a vague verification loop.
Sure Sol 6.1 on low/medium can be considered "cheap" considering its abilities, but it's clearly not meant for the same type of tasks as Luna. You just grow an intuition for when something is solvable with a Luna-sized model. I'm guessing Terra is good for the tasks that you'd typically use Sol 6.1-low for, or for Luna type tasks (or other not-novel problem) where you want to reduce the risks of suboptimal decision making. It's exactly what I'd go with for feature additions where the correct solution is already mostly designed/specced-out.
The way to think of this is that (current) inferencing hardware has natural cut points of efficient scale (assuming some optimal architecture which at this point is almost always an MOE variant), so you always have an incentive to provide 3-4 model sizes (which nicely enough often means diff speed to answer also).
You are also rolling the training of your models by size and the lat est small model can sometimes beat an older model.
For a while we thought you just train the big model first then distill smaller models from it but since we've learnt the importance of on policy RL (basically the way a small model figures out a problem is very diff from a big model so it's best to RL optimize each model distinctly)
Name means size which means price
Yes, a more modern smaller model can beat an older bigger model at some tasks, that's what can be confusing.
> Did Microsoft's chief branding officer move over to OpenAI?
Obviously not, otherwise the four models would be called Copilot, as well as the harness.
You're just not paying attention. If you run copilot in copilot using copilot effort, it performs much better than copilot in copilot using copilot effort. This is clearly explained if you visit copilot.com using copilot and then search for copilot.
“Introducing our first mobile devices: Pocket Copilot Luna for on-the-go inference, and our new tablets, Copilot Sol+ and Copilot Astra Duo.”
Training models is still not a precise science.
If you scale a model down, or switch out attention with linear attention, or introduce KV-compression, etc., you don't really know how it will affect model performance before you've trained the model.
And once you've spent millions training it, you might as well make it available if it's pareto optimal for some tasks. It also gives you more insight seeing how people use it in practice.
Flip it around. The number (generation) is the most important. Just like it's often with hardware.
And some variants (e.g., the medium one) being skipped or delayed for a certain generation doesn't really change the overall size branding.
Agree, it's confusing. I hear people discussing which model to use for what tasks and my head starts to spin because there are so many properties to account for and it's generally hard to reason about LLMs.
"Speed, quality, price: pick two". I think I'd rather see a choice like that than the current convolution of $-per-token, sub quota, latency, simulated reasoning level, model size, context size, etc, etc.
>But then things go confusing, they released Sol-6.1 which sometimes is superior to Astra, and discontinued Terra which made it look like this
I don't want to rush to their defense here, but, when model families are getting refreshed every 6 weeks, there is going to be significant overlap, particularly with these staggered releases
And low/medium/high/xhigh/ultra reasoning. And pro mode and ultra-speed versions. There's at least 25 options to choose from.
Next AGI test, when model companies have consistently good version branding.
The moon is smaller than the earth. The earth is smaller than the sun. The sun is smaller than all of outer space.
Wait I always thought terra was the smaller model and the further away you went in space the bigger the model (luna is further away so it’s better than terra)
It's especially confusing because Sol 6.1 is mostly just Astra but 10x slower at 1/10th the price - unless you pay for it to be as fast as Astra - at which point it's the same speed and price...
It's nice that they have Terra which seems to fill a void that's been needed since Sol 6 focused on cost (at the expense of quality), Sol 6.1 got a huge boost in quality, but became unusably slow unless you paid Astra prices - at which point, why not just use Astra?
Now at least we have a viable other option again.
Huh? I have not observed Sol 6.1 and Astra 6 to be the same.
the only confusing bit was dropping terra for a bit,
Luna (Small), Terra (Medium), Sol (Big), Astra (Bigger)
is perfectly clear, sol beats astra sometimes because Astra 6.1 was scrapped. and note that opus 5.5 beats fable 5.1 most of the time too without a skipped model.
You're not alone on this. They have too many models, and too many effort levels, it starts to feel like what to chose.
Different classes and different generations. Isn't it roughly the same as what car manufacturers do? Or even Macbooks? 2014 macbook pro has worse performance than 2026 macbook air, this doesn't confuse consumers.
No idea, i still try to figure out how to use the right one per task.
I suspect 6.1 Astra is the model that kept showing up in the news stories about an internal unreleased model trying to hack the Pentagon or whatever.
also am I the only person who remembers Do Kwon and the Terra/Luna debacle?? how about Gaia/Selene (admittedly harder to non-native english speakers)
Bet on purpose
“Test them all”
Pelicans for GPT-6.1 Terra: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Interesting that both "none" and "low" reasoning levels appeared not to spend any reasoning tokens at all, and in fact low ended up using less tokens and costing less than none here.
"medium" did use 496 reasoning tokens, but that's apparently not enough here to return a visible reasoning trace.
OpenAI's current frontier model lineup is gpt-6-astra, gpt-6.1-sol, gpt-6.1-terra, gpt-6-luna, so here's a comparison grid of all of their pelicans at different reasoning levels: https://static.simonwillison.net/static/2026/openai-pelicans...
What's the point of using it without reasoning? Isn't that what Luna's for?
It's neat that we can now compare the Terra without reasoning pelican (cost 0.64 cents) with the equivalently priced Luna with reasoning pelicans (Luna xhigh is 0.40 cents, Luna max is 0.91 cents). I think the Terra low one compares favorably to those, though I'd suggest also comparing them on tasks more representative of real work than generating pelican SVGs!
a bigger model still has more built-in knowledge to draw from, even without reasoning.
I appreciate that most of these pelicans are not holding the handlebars, i think that's more realistic than the alternative
I can literally see what the training data sample looked like that they included, so it performs well on this benchmark.
12 seconds for a pelican on low. Pretty good.
Do you have an agent scanning the front page waiting to post the pelicans?
I should do but no, it's just a bad HN habit.
Seems Gemini is way ahead in pelican drawing
Apparently Google made it a point to train their models to output SVG, among other things, so that helps
Can anyone give me the compelling reason for Terra and Sonnet to exist?
Virtually every task I give to either is better served by going up a tier and down a reasoning level, or down a tier and up a reasoning level. I simply don't understand their use case. No testing I've done has shown them to have a niche where there isn't a cheaper-with-equal-results or more-performant-at-same-cost option.
I think for the majority of the audience here who use these models inside of coding agent harnesses w/ subsidized monthly subscriptions, there likely isn't one. Depending on how Terra contributes to the rate limits, it may make sense to use it for some subset of a repeated workflow to get more mileage out of sub.
If you pay API prices for inference in your product stack & have specific use-cases w/ custom evals, these smaller models provide an additional cost/speed lever to adjust. Ultimately, I think the point is to compete on cost for use-cases where Sol on the lowest reasoning is overkill.
If you have a fairly straightforward usecase where Terra performs just as well as Sol, but is faster and cheaper, why not use it?
Yes but you are missing the point. It's the case where Luna is not good enough, and Sol is too much. The problem is that on benchmarks Luna xhigh is typically comparable to Sol low in terms of capability. So where does Terra fit in?
Idk… I’ve written 100k sloc in ~2 weeks on the Claude £20 subscription with sonnet medium alone. Seems pretty good to me.
On the API we use a lot of Luna (cost efficient) and sometimes Terra level models. I think we never use a larger model. These are models that our app runtimes depend on.
On subscription for coding and stuff I am basically on Astra Ultra all day long.
They are the biggest models supported in Github Copilot Pro
https://docs.github.com/en/copilot/reference/ai-models/suppo...
You should look at model reasoning level curves on the cost/performance chart on artificial analysis. If you compare the curves for haiku/sonnet/opus I would agree with you that sonnet serves no purpose because it is priced too high. If you look at luna/terra/sol you will see that terra is priced such that it successfully fills a gap between luna and sol. And when you add speed into the equation it adds another dimension where terra makes sense. terra low falls between luna xhigh and max on both cost and intelligence, but it is way faster because it uses way fewer reasoning tokens. So I would say there are workloads where terra makes sense. But from my analysis I don't see any workloads where sonnet makes sense.
Sonnet is way faster than Opus when doing everyday tasks, I think on the subscription plans it's a better option when you need something done faster than Opus' long thinking sessions.
There is a chasm of performance difference between haiku and sonnet ime. Sonnet is cheaper and faster than Opus, so if you're providing a service via api where cost and speed is important, it's a big difference.
Except Sonnet is slower, more expensive, and worse at reasoning levels above ~medium. Opus on low is faster, cheaper, and better at doing things.
1. In Anthropic's benchmarks, Opus's advantage over Sonnet diminishes at high reasoning levels. Sonnet max actually beats Opus max.
2. There are rote or simple tasks where higher reasoning isn't needed and one model vs the other wouldn't even be distinguishable.
3. How many tokens are used during reasoning depends as much on the task as the reasoning level.
4. Sonnet is faster and half the price per token, so Opus will only be cheaper if it can spend fewer tokens. This won't be the case for all tasks. Opus and Sonnet will also consume the same number of tokens to load context, so the relative price comparison depends a lot on where the task is context-heavy or reasoning-heavy.
> Sonnet max actually beats Opus max.
What are you smoking and can I have some?
For coding, no.
I have some business automation process for some customers that I maintain, where Terra with low reasoning ended up with the best price/performance & latency. With an equivalent Luna with high reasoning, price was usually better, but quality had a higher variance (very much depends on where the thinking takes the model) and latency for high reasoning was worse.
If you trust these benchmarks, I'd argue that there's almost never a reason to use anything other than Terra 6.1 xhigh.
If you use API in agent or your own integration with software, such models make a lot of sense. We use Terra 5.6 for "question to SQL" translation for example. It's good enough and affordable enough. Luna will probably not handle complex queries well while Sol is too expensive.
I've been refused to do cybersecurity (read-only investigative access against the recent brotli cve) with Opus 5.5 when I had full SSH access, so I guess you've got to have a fallback?
i use sonnet to generate answers to help-desk tickets, which it does for significantly cheaper and insignificantly worse than opus.
if you're just talking about use in codex or claude code, then sure, leave it on opus all the time. but for running routine api requests, sonnet is great.
I cannot figure it out either. I use Astra Ultra to plan and 6.1Sol High or X-High to build and even on huge projects with 10s of 1000s of pages it is fast and competent
Sometimes this tier is faster, and getting a good enough result in a few seconds instead of a better result in ~15-30 seconds is a meaningful result.
I think it's also a fake lever to give to finops / it teams to try and keep costs down.
> Can anyone give me the compelling reason for Terra and Sonnet to exist?
To justify the pricing of Sol and Astra. Old business trick.
This is the only reason that makes sense to me. Thank you
I'm on Claude's default plan for personal use, and Sonnet is plenty enough for small projects while not hitting my usage limits too hard.
Cheap and affordable. Dispatching proper tasks to these models can save you a huge % of your quota. Especially if you _know_ the task is simple (ie add something here or there, or archive a bunch of files and reformat kinda thing)
Maybe it's my tasks, but in the cases you're describing, I'm usually better off going down to Luna. I specifically don't see the gap where "It's too complex for Luna-high" and "It's not complex enough for Sol-low" exists.
I've personally found that most models completely collapse at sub-medium effort. I generally run most of my stuff at medium/xhigh with heavy core tasks @ max
Similar 75% price cut trajectory as GPT-5.6 Sol -> GPT-6.1 Sol. Overall, Terra is still half the price of Sol.
But like the original release of Terra, it's still a weird middle child between the small-model-that-got-substantially-better (Luna) and the large-model-that-got-significiantly-cheaper (Sol). The TerminalBench 4.0 cost-per-task results are a bit confusing and I'm not sure why you would pick Terra over Sol in that case. Overall this release seems like a flex on Sonnet 5.5: a model which is having its own middle child problems between the godly Opus 5.5 and the surprisingly capable Haiku 5.5.
the surprisingly capable Haiku 5.5
Destroyed by Anthropics crazy decision to price it x5 after 100k context.
I don't think software devs are the audience for 6.1 Terra, but instead computer work, automation, and technical writing.
Honestly I'd much rather take this model then.
It is having improvements in other fields than coding while being more cost efficient.
If all I want is coding at a cheaper price, then deepseek is a good model for that.
I am more than okay to have a bifurcation of models good for coding and models good for everything other than coding. I strangely imagined it many years ago.
Everyone is suggesting OpenAI to release a model close to opus 5.5 yet it's releasing this. On one hand we can imagine that maybe their training cycle wasn't able to produce this whereas gpt 6.1 terra was made.
But on the other hand, I also now wish to see an OpenAI model which is good for just pure coding, like an absolute beast at it while being cheap (so the opposite of terra)
ironically, one of the few appreciations that I might give to OpenAI if they pushed this model to prod and are facing some pushback as developers like us want/only see coding models. They could've benchmaxed it and it would've been slightly irrelevant. I'd much rather see reality.
At the very least, this new model won't make things more expensive for anybody and would only make use-cases cheaper depending on what they are, so it's a fine deal for me.
I also wonder maybe this is why open source labs aren't targeting this space and why the only focus of AI labs is coding nowadays. Everything else just doesn't seem so worth it from investor, even end user, training perspective.
Maybe this is the reason why we have gotten almost an exponential growth in coding capabilities as that became the most (if not only) capability that truly really matters.
To not wish for discussion to go to a tangent, nonetheless this could also explain why our work as a software engineer is going through a slight crisis of meaning as these AI models have gotten impressive at coding.
As an (aspiring) developer (currently in 1st year college), I don't wish for AI to replace coding even more for me or to push it even more or to suck the joy out of it even further. So I am more than happy for these models to stay as they are and rather make advancements elseplace. Ironic because it is us developers who use AI as well and we want cheaper workflows and so we would also want models to be cheaper and better at coding. There are some fine lines of contradictions and irony in all of it. I'd much rather also have open-weight models of such terra-like models but I imagine that because of the reasons that I mentioned above, we are most likely gonna see it the most focused on coding but still open-weights is pretty close.
The more choices the better (though choice-fatigue is a real thing), I hope I was able to explain some of the nuance I feel about this particular launch.
Well put.
(For some reason they say "Cache read" instead of "Cached input" for 6.1 Terra)
6.1 Astra and 6.1 Luna must be coming up soon. Will they reduce their prices further?
I guess 5x/95% is where we're landing with the 6.1 family, that's nice.
I don't understand, in my tests, on max effort, it's worse than 6.1 sol and 2x more expensive[0].
Used on medium seems to be the best use-case[1] and cheaper than 6.1 Sol.
[0]: https://aibenchy.com/compare/openai-gpt-6-1-sol-max/openai-g...
[1]: https://aibenchy.com/compare/openai-gpt-6-1-sol-medium/opena...
It is fast, but more verbose (a lot more verbose than Sol at higher efforts)
I rather enjoy OpenAI's focus on cost-per-task.
I'm optimistic about them tackling frontier intelligence next. Astra 6.1 please?
They're doing its because they recently cut plan usage by about 50%, leaving Anthropic providing about 5x the value on their plans. [0]
The "cost per task" metric is just a weaselly way to sound like they didn't just double the cost for most users, and aren't having their proverbial lunch money stolen. They can no longer claim to provide more tokens, more capable models, or more usage in general. So they made up a new metric where "task" could be defined to mean whatever they need it to mean.
They had to become VERY specific in their wording to present any kind of value proposition that will entice people.
[0] https://newsletter.semianalysis.com/p/anthropic-subscription...
> leaving Anthropic providing about 5x the value on their plans
5x is calculated in terms of API costs, nobody pays those extreme Anthropic prices outside enterprise users.
> like they didn't just double the cost for most users
Most users aren't on their $200 plan. Most users (including on the $200 plan) continue being subsidized.
> 5x is calculated in terms of API costs, nobody pays those extreme Anthropic prices outside enterprise users.
Ignore value then. Anthropic still provides about 3x as many raw tokens. The 20x\200 plans provide about 10B for openai versus almost 30B for Anthropic.
Do you really think that Opus uses 3x as many tokens as Sol per task? I don't.
> Most users aren't on their $200 plan.
Fair enough. I should have said most heavy users.
> Most users (including on the $200 plan) continue being subsidized.
They continue to be fifty percent less subsidized, I guess.
Look, I don't fault them for trying to become profitable. I fault them for the weird weaselly business speak about cost per task. It's BS game playing and we all know it.
I'm honestly rooting for them, and hope it all works out. For now though I'm pretty salty that my bill jumped by 150% in one month. Gaslighting about made up metrics just makes me more angry.
People in HN live in a bubble. Plenty of real world use cannot use subscription plans (not even talking about enterprise agentic use) and is absolutely at the mercy of the cost per task metric.
We agonize over cost per task to the extreme for services/functionality built atop of APIs and can absolutely guarantee you that it is the real world metric that counts most right after “is the model good enough for the task at hand,” and furthermore that, yes, in fact, it is quite often that we see certain models/providers easily using 3x or more tokens on average for the same tasks (or even the exact same task) than a supposedly similarly matched model from another provider. Once you become intimately familiar with the thinking traces from specific models, you can preempt this to some extent with manipulations to the system prompt (essentially heading off entire thinking branches) but the nuking of reasoning traces in the name of preventing against distillation attacks has made this harder.
Cost per task, time per task, quality (good luck measuring objectively between models) are the metric that matter. I don't care how much a token costs, I'm not buying tokens, I'm buying solutions.
Exactly! I called that the Operations Equation and wrote a tiny bit more here about it: https://zero2data.substack.com/p/the-operations-equation
Once you are inserting models into non-creative (and software development is creative work) it is imperative to start creating your own small evals. Not just to check how different models perform in that equation, but to catch model drift over time.
the point is that "task" is a dubiously defined unit that the frontier labs adjust to whatever makes them look good.
Are OpenAI's models delivering 50% (or more) completed tasks than they were last month? Are they delivering 50% more than Anthropic per token?
Nope to both. Not even close.
The only way it works is if you define "task" in a way that OpenAI is good at it and Anthropic isn't.
I kind of expect them to go right to Astra 6.2, 6.1 was cancelled over the cybersecurity hubbub. I eagerly await either one.
OpenAI repeatedly keeps hurting user workflows by altering the model generations family for each new version in unpredictable ways. As such, it is not a sane ecosystem to develop in.
If you are confused, that is how this company was always been, remember GPT 4, then they did some weird versioning.
I personally like Terra, it kind of does what I would ask Sol or Astra, but dont have to count tokens as much. And Luna has been amazing tbh, it takes designs from prototype and in record time is implementing it, so that was a huge win.
But yeah, they always had confusing way of naming products and if you are not following news, you will be confused about what it is.
Wow. If those charts are true, this effectively matches GPT-6-Astra at a fraction of the cost, and is neck-and-neck with 6.1-Sol.
Looks great on the first chart, but terrible on the Terminal Bench 4.0 where it needs xhigh to come close to 6.1 Sol, and costs more for a slightly worse result.
I would guess only on the nitpicked benchmarks?
The pelicans place it nicely between Sol and Luna in terms of both pricing and results.
pelican is also nitpicked (and probably the most useless)
Cherry-picked is the term, not nitpicked.
Yes, it may be dumb, but it isn't useless because it is both a) is a benchmark and b) is not picked by openai
IMO it's useless because we don't even know what a pelican on a bike is supposed to look like, It's very impressive that that models can make these directly using SVG, but we are way past the "wow" stage and test that can't be gamed and are easy to prove how good the model is in that specific arena are hard to find.
The only thing that pelican have is "models where not trained on this" maybe that true months (or years ago) now it's on every HN post and these AI labs are monitoring HN
Not necessarily, they may have selected benchmarks where Terra performs relatively well compared to those models. Pretty much every model release these days just cherry-picks & highlights the benchmarks where their latest model performs best. We'll have to wait for the complete benchmark suite to have a better picture.
the only things we've got to actually make a judgement on are vibes and token anxiety. vibes generally indicate that more expensive == better, but more expensive == more token anxiety. so you end up using the most expensive model x effort that keeps your anxiety in check. yeah, maybe terra is great at technical documents. i'm not going to use it because i'd need to spend a couple of days writing technical documents to work out if it was better than sol and then what if it isn't?
is opus 5.5 better than sol 6.1? hmm, i really can't tell. maybe claude is a bit more diligent and identifies more issues up front. but sol has this can do attitude and the codex harness just gets on and does it's job done.
Why do Anthropic and OpenAI think it necessary to split their model into 4 versions (Opus, Sonnet, Fable, Haiku, Astra, Sol, Luna, Terra) vs. Google's 2 (Argon, Flash)?
Google has flash-lite, too. I liked OpenAi's scheme better, as long as it is parallel and makes sense.
cool update. Like super cool.
Just one thing about the advert, how can I be the only person here thinking the ad is weird - like for the marketing dep, its a bit weird the ad features the same tram that broke and tragically like ended some people - happened for like the first time ever in 100 years, like a year ago ( 14 months ago) just a bit weird.
Im honestly not trying to be negative, but its kinda weird I kinda feel I should mention it because noone else has?
Im sure literally anyone from southern europe must have thought this the second they saw the ad. It was big news last year.
Am I mad? Just a heads up for PR to always check things like this first when releasing a video? Maybe its just me? Just trying to help
It's a subliminal warning. "On the trolley problem, this model will try to kill you and make it look like an 'accident'."
Still can't really forgive any of them for essentially stealing most of this knowledge regardless...
And then talking about other people distilling from them...
I feel like OpenAI hasn’t been as (publicly) hypocritical and petty about their distillation spats as Anthropic has been?
They are public about it - see their usage policy: https://openai.com/policies/usage-policies/
e.g.: "You must not use OpenAI models or their outputs to generate, prepare, select, or supply training material or signals for models that are intended as alternatives to or substitutes for OpenAI models or services".
Anyone who has data that is so important to their livelihood should probably make backups.
The results for factual error rate on "difficult" prompts is staggering to me. Particularly the difference between 5.6-Terra and 6.1-Terra. These are wildly different Pareto fronts. Our custom enterprise agent is currently running on 5.6-Terra and it's already very good. I look forward to updating the configuration.
I'll wait for the runebench results!
On AA intelligence index, this fills in the cost per task frontier nicely between 6 Luna and 6.1 sol. So when 6.1 Luna inevitably releases...
The OAI model tango seems to not have an end in sight, for better or worse. Give me more API toys!
This tier is what I'd actually point opencode at for PR reviews. The top models are too slow and expensive to run on every diff.
These benchmarks made me want to try sonnet 5.5
These "Pareto frontier" performance vs price graphs have become very popular and I find them mostly uninformative. Pricing is subjective. You chose the position on the X axis, and which measures to use on the Y Axis.
Am I reading the Agents’ Last Exam v1 graph correctly where it's saying GPT Luna Medium performs better than High, but XHigh is better than Medium?
The error bars on eval graphs are commonly like ±5%. It's fairly common to get out-of-order points when the gap between them is only a few percentage points. If we ran it again, or ran it with minor alterations, I'd expect the ordering to be expected. In general, we don't massage any of our eval numbers and for good statistical hygiene we'll report the first run we do, even if later runs show slightly smoother curves.
(I work at OpenAI.)
That is correct. The gpt-6 models are strange.
I am wondering if there is a site with a perf/efficiency chart of the models over time? I am wondering is the pace still speeding up?
Sell it to me. Why would I ever give money for this?
You are paying API prices, and 6.1 Terra gets you 90% of what 6.1 Sol gets you at half the price.
My understanding is API prices are more expensive then monthly plan, is this true?
My plan for OpenAI recently expired and now I’m thinking to test Anthropic ecosystem. My primary use case is programming only.
Should I go with API use or subscription? For coding harness, I jump between codex, OpenRouter, pi, whatever coding harness is trending on that day
Individuals are not paying API prices for coding. Everyone uses a subscription.
Anthropic does not allow you to use 3rd party harnesses -- you have to use Claude Code. Claude Code is an excellent harness, though.
If you want the freedom to move between harnesses you should use OpenAI or xAI
Subscriptions provide thousands of dollars' worth of tokens if measured at API prices. While the value was once absurd, it is now slightly lower but still much cheaper.
anthropic subscription does not allow 3rd party harnesses i believe
not gonna lie they did really bad and this drained my usage SO FAST. that model used lots of unnecessary skills and read through way more memory files needed.
It literally burned through 2 subs.
They say it has "A significant step up on factuality and coding". My browser is underlining "factuality" in red. AI hallucinations or hallucinatiality are still a big problem in my experience.
First known use of factuality was in 1859. It's a "real" word. You can look it up.
https://www.merriam-webster.com/dictionary/factuality ?
GPT-6 only just came out! The AI landscape is always on the move!
2026-09-03 is almost 2 months ago already - that's basically half a lifetime in LLM land!
True!
https://artificialanalysis.ai/#intelligence-efficiency-tabs
According to this, Sol 6.1 x-high is both cheaper and smarter than Terra 6.1 max.
i'm not sure anyone cares about Terra though.
Previous Terra model was so good it was discontinued.
OpenAi needs a model to compete with Opus 5.5, and that's not Terra 6.1.
It could have been Astra 6.1 but they don't have the courage to make it happen.
Is it stronger and cheaper than DeepSeek V4.1 Flash 1017?
If not, yawn...
AA has limited 1017 benchmarks so some terras versus 1017 Max: max - 3 points smarter, 50% more expensive xhigh - same score, 25% cheaper high - 1 point dumber, 50% cheaper
Do people use deepseek for real world development or just for building experimental large projects?
Yes, V4.1 Flash 1017 is my daily driver cloud model. Copes fine with what I'd consider difficult brownfield development, like fixing bugs in vLLM prompt cache and coming up with small/sane overlay patches for them. Maybe it's disappointing for people who want to one-shot a game or website from a single prompt; I haven't tried that.
It's nice that their peak-rate hours end quite early in the day UK time.
I use it at work. It's dirt cheap, blazing fast, and is more than enough for most coding work.
If you live in a country with a weak currency, DeepSeek models are definitely a godsend.
You can do real world development with weaker models than Deepseek Flash.
I still use SOTA for the main model though, but also I'm perfectly aware it's mostly out of FOMO.
Deepseek is more than enough for default subagents though. Any well defined and verifiable task and there's absolutely zero difference between SOTA and good-enough models. Used to be just speed, but Deepseek is fast.
It's only a little better than Deepseek with Max effort in benchmarks. Doesn't bode too well for 6.1 Luna.
The benchmarks are already saturated and the curves for Terra stick to the axes.
Really hate how they only show it vs Opus @ Max and not Low/Medium/High.
Underwhelming
That terminal bench 4.0 result is... bad? Terra XHigh more expensive than Sol Low and still does worse.
It's very likely this is a next checkpoint of GPT-6 "Sol", over-optimised for the web chat and offered to Codex subscribers as a "please go away" offering.
If this was half the price, I'd be very interested, but at 1/5/0.05? Feels like another panic release because they can't get Astra 6.1 out of the door...
Scoring "bad" on Terminal Bench 4.0 can also be a sign that it's cheating less: https://preferencemodel.com/blog/terminal-bench-4-is-saturat...
People keep asking OpenAI for a mode that's as good as Opus 5.5, and they release this instead. Interesting strategy. Let's see how it works out for them.
They don't have anything competitive with Opus 5.5 yet. There is a supposed model in the works, "Bel" last time I paid attention to gossip, that is thought to be better, later this year. Allegedly Anthropic caught OpenAI off guard and / or their different training approaches ended up working better than anticipated.
My expectation is Fable 5.5 is next overall improvement, and my bet is OpenAI's "Bel" will be the next competitive one (or, their lack of safety culture will bite them/us in the process).
There’s also a chance they surrender the dev market altogether to Ant and just focus on the “normie” segment with dots, computer use, excel etc. In which case they won’t have anything competitive to opus-5.5 anytime soon for SWE.
Luna and Sol-6.1 are crazy cost effective and all flows here have switched to them, so I'd say really well so far.
There's very few stuff that needs Astra/Opus level intelligence.
Luna 6.1 would be amazing.
This release is interesting for enterprise at API rates.
At least unless the Terminal Bench 4.0 score is more indicative than the other ones, because that one shows it clearly worse than 6.1 Sol for both performance and cost.
Taking the benchmarks in the article at face value, this is ~on par with Opus 5.5 for ~1/4 the price!
Which is why I don't take the benchmarks at face value...
So you have decided the model is bad without having used it yet?
If they had a model on par with Opus 5.5 there would be no need to undercut it by that much, that's probably the logic.
Anthropic models are comically overpriced. They also gobble up an obscene # of tokens to get the same result as others. The price delta is literally 5x.
No, I have decided I don't trust benchmarks without using the model first.
Taking the benchmarks (as reported by OpenAI) at face value, why wouldn't they show Opus 5.5 on the OS world benchmark, and show no Anthropic models at all on the "factual error rate" benchmark
...no it's not? At max effort, it scores 50% on terminal bench vs. Opus 5.5's 70%.
ALE (top benchmark in post) is 60% for Terra vs 63% for Opus - I guess we'll have to wait for AA Index!
48 for GPT-6.1 Terra Max vs 58 for Opus 5.5 Max on AA Index
I hate Opus 5.5, constant overreach and over-engineering for any project
That's wild to say when comparing to the OpenAI family of models. I've never seen so many tests that validate tests that validate tests.
Right?! I think sol has some obsessive tendencies, i literally caught it the other day spawning a sub sub agent doing a final fourth test for a piece of code that isn’t even critical say in embedded or similar, now I have to instruct it to only test once!
Medium or High? Depending on the task, you might want to switch between them.
You’re holding it wrong.
"better writing"
let's see if it can pass the Pangram test. Many services purport to fool it, but fail.
Fooling Pangram and writing well are not the same axes, even if there’s some correlation.
Everyone defaults to checking pangram when AI is susepcted, so it's still an important test.
Because Pangram is snake oil.
2025 called and it wants it's take on AI detection back.
https://www.pangram.com/blog/third-party-pangram-evals
Note that while that page is from Pangram, the studies on it are not. If they are excluding studies that show Pangram performs poorly in realistic conditions, please share those studies!
If Pangram was snake oil (it's not but could be more mindfully marketed), it's definitely not now that every major lab has added watermarking.
You wouldn't even need to be able to read the information the watermark encodes to recognize it exists in a piece of text
This is incorrect. You cannot identify the presence of text watermarks if you do not have access to the models so that you can spot where expected tokens diverge from actual tokens based off the weights. That does not preclude Pangram from being able to reliably identify LLM-generated text from models whose outputs they’ve trained their own detection model upon.
They have unwatermarked text from the same base models, human text, and watermarked text.
If they couldn't train a reliable detector with that they wouldn't exist.
If they had any courage, they would release 6.1 Astra, but they don't.
Update: Look at the OpenAI employees downvoting me.
The next model will likely be Nova