An economically successful step would be to improve training cost/frontier advance.
They essentially failed at that, assuming they kept their training costs for the model step at least constant.
Notable takeaways:
- Sol-low - cost-competitive with Haiku, while outperforming Sonnet on medium
- Sol-High - Competes with Astra medium and beats Opus 5.5-medium, at 20% the cost
- Sol-Max - Beats Opus 5.5 high, Astra high, and competes with Opus xhigh, at cheaper than Sonnet and a fraction of the Opus cost
And when you look at time spent, With the exception of Sol-max you get better results in less time for less money.
It's rather hard to look at opus-- or any of anthropic's models-- and say "economical".
And for most people here who care about coding performance Anthropic has a substantial lead[1] - although sadly they don't have coding benchmarks separately for Opus 5.5 Medium which I think it the sweet spot.
The cost tab really should have something like "how many tasks can you do in a particular subscription tier" for many people. I realize this changes often though.
[1] https://artificialanalysis.ai/?models=gpt-6-astra-high%2Cgpt...
It seems like most analysis is basically "how many tokens do you get with a subscription and how much would those tokens cost at API rates". Looking at it that way, you do get more dollars of API equivalent usage for your subscription.
But Anthropic's API pricing is a lot higher than OpenAI's (Opus 5.5 is 2x of GPT-6.1 Sol). The number of tokens consumed per completed task is also far higher than GPT-6. So when you look at the number of tasks you can complete within a subscription, OpenAI looks like a much better value.
That is before you consider that OpenAI subscriptions (other than the cheapest one) don't have 5-hour usage limits, and also generally give you multiple free resets per month.
There was some chatter on Twitter, Reddit and here when an earlier version of this came out where people normalized it by completed AA or Epoch task and Claude models still came out well ahead.
In practice OpenAI subscriptions are more competitive than they appear because OpenAI does a lot more subscription resets. But that's hard to account for!
Per task, Opus 5.5 costs ~6x of Sol 6.1. Then you add on the lack of a 5-hour limit, no limit on which models (like Fable's 50% of usage limit) you can use, very frequent resets, and that OpenAI doubled the usage limits after that analysis was performed.
Even though the Claude subscription is equivalent to 5x the $ value of API priced tokens, it seems like Codex is the better value for getting things done.
https://artificialanalysis.ai/?models=claude-fable-5-2-high%...
I don't think so? I think OpenAI had planned to half the limits but decided not to go ahead with it, but this analysis was done at the limits we have now. I could be wrong about this though?
> Opus 5.5 costs ~6x of Sol 6.1.
My point in part is that you can't compare like this.
Opus 5.5 is a more capable coding model than Sol 6.1. AA benchmarks show Opus 5.5 max effort scoring 62 vs Sol 6.1 max effort on 54: https://artificialanalysis.ai/agents/coding-agents?agents=de...
Unfortunately AA doesn't report non-Max efforts for their coding benchmarks.
But no one runs them on max. For Opus the advice is to usually run it on medium effort, and I've found that advice to be good.
For me to get equivalent results on Sol I need to run it on high effort, and I need to use Astra as an orchestrator (whereas Opus medium does great as an orchestrator too).
That combination pushes the blended token pricing up a lot.
I can't give you hard numbers on this, but at the moment my Codex credits get used up a lot quicker than my Claude credits on similar tasks in the same codebase.
I'm glad we have two great models competing, but at the moment I'd say Opus has the edge.
I think it is a narrow view to only consider the AA coding agent index as the only metric, especially with the models running at max.
For the link you shared, Opus used 3.7x as many tokens, and 2.5x the time to get its score. In the real-world building things with these tools, that kind of increase in time and cost (or subscription usage) matters, at least to me.
When comparing at more reasonable medium or high levels of reasoning, the gap between Opus and Sol narrows considerably. Opus being about 2 higher than Sol at the same reasoning level. However, Sol 6.1 xhigh matches Opus 5.5 high on scores (52), but while still being slightly faster and costing 78% less ($0.76 versus $3.46).
https://artificialanalysis.ai/models?models=claude-opus-5-5-...
> https://artificialanalysis.ai/models?models=claude-opus-5-5-...
Yes, but my point is that these are non-coding benchmarks at these non-max levels.
On coding at non-max Opus really does seem noticeably better, and my subscription seems to go longer. Look at the other comments in this thread - every single one says the same thing.
The difference is the Opus API cost is $1200, and Sol charges $50, and finished in about 6 hours less.
Is the Opus airport better? Maybe, possibly, but I'm a lot less annoyed with Sol because I can correct course a lot quicker.
Plus Sol's prose is substantially less obnoxious.
If you leave xhigh and max off.. you still get better performance than Opus 5.5 max
Opus 5.5 Max - score 59 - cost per task - $10.04
Fable 5.2 High - score 62 - cost per task - $9.08
> They essentially failed at that, assuming they kept their training costs for the model step at least constant.
What? No!
This is completely wrong!
This proves that inference scaling continues to work beyond the previous frontier.
Claiming that higher inference costs shows failure is an odd thing to think. We've seen this optimization cycle over and over, most recently with Opus 5 -> Opus 5.5
Opus 5 was slow and expensive. Opus 5.5 came out and was better, fast and much cheaper for them to run.
There is nothing to indicate the same can't happen here.
Anthropic probably has multiple options open to them with this model, and they'll probably explore all of them:
- Use collected Fable 5.2 traces to improve reasoning economy for Fable 5.3
- Use Fable 5.2 as a teacher model for Opus 5.6
- Add additional fields RL data to expand Fable capabilities for Fable 5.4
Same is happening in LLMs. Sadly, the Golden Era is over.
And maybe that's how all of it is going to evolve in the future – so you pay per solution, not tokens?
Perhaps, it's a bit like being a super smart person in contemporary society: at some point, being smarter doesn't really reliably translate into a better life.
When you learn from available human data, those text books max out at some point. You cannot learn what isn't there.
To do that you need a larger model to have many many branches to try.
This is a dataset basically constructed by taking a bunch of DeepSeek R1 trajectories and filtering out the incredibly long rambling ones: https://huggingface.co/datasets/open-thoughts/OpenThoughts-1...
There's also a bunch of RL research around strategies to coax the student model towards more concise trajectories.
In general, a lot of people in industry are finding cost savings these days by tuning a smaller model over longer agent trajectories (not exactly the same as distilling the reasoning traces, since those aren't exposed by many frontier labs anymore) collected from their production agents. Typically works best when you have a pretty workflow-y agent whose tasks all share a similar shape.
Let say you have a model that score 100 in a benchmark and you distill it into a model half the size. It has very few guarantees that it will score close to 100 again.
But I assume one of the hard problems is to know that the model might not perform well enough outside of the distillation part.
There is also another caveat that models do display some level of sub language control, by this I mean that innocuous text might contain information relevant for the own agent or for other agent with it shares tasks. (Like when your partner says there is nothing wrong) Because distillation uses very few resources, we don't know exactly what gets send forward or not.
As Caleb said, many companies found that they can tune small open weight agents over procedural known tasks because that's easy to inspect and check and because there is little space dor ai-loitering, it tends to be less problematic. The key difference, at least in my understanding, is that you are not concerned with the weights or reducing data. You are basically tuning agents existing weights to emulate another agent answer.
They are doing IPO now because it might become apparent that they reached a limit in the next year
And I am simply ignoring the ridiculous comments around local model usage from certain CEOs.
Having these highly rated models reflects (potentially) the frontier of what they are capable of producing. Essentially its and advertisement and shows directionally where the smaller faster model's capabilities are headed.
The "wall" is in increasing intelligence per cost. No one is saying the absolute intelligence won't continue increasing, it's that the rate of intelligence increase per cost is falling from super-linear growth rates to linear at the same time the companies are backing out the investor-funded subsidies they were using to buy adoption and user growth numbers. You're correct that there is a 'trickle down' effect to lower tier models but OAI/Ant are clearly using this to segment customers by willingness to pay and their IPOs depend on extracting the maximum amount possible from each segment.
That's why the prices of frontier models have been skyrocketing, though they are using every trick imaginable to mask this being clear and obvious. They take with price increases, token deflation, shifting plan levels and limits but muddy things more by giving back some (but always eventually less) with resets, subs, etc.
They absolutely have said that.
> That's why the prices of frontier models have been skyrocketing,
They haven't.
"The cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year."
"The cost of a given level of performance often falls fastest right after that level is first achieved, that is, when it is state of the art (SOTA). We see this pattern on three of our five main benchmarks of AI capability. Averaging across all five, cost falls 66% per quarter (75× per year) for performance that has just debuted as SOTA."
https://epoch.ai/publications/the-plunging-price-of-thought#...
On a per-token basis Fable 5.2 is cheaper than 5.1:
5.2: $8 per million input tokens and $40 per million output tokens.
5.1: $10 per million input tokens and $50 per million output tokens.
Compared to older (and much less capable) frontier models it is also cheaper. For example GPT-4 was $30 per million input tokens and $60 per million output tokens
https://developers.openai.com/api/docs/models/gpt-4
Be very careful comparing the Artificial Analysis "cost per task" scores. Fable 5.2 is expensive there because it is doing tasks the other models fail at.
Notably, in order to achieve a specific quality, the model has to "reason" for every prompt proportional to the training inference.
This is arguably the case of diminishing returns being reached, as answer token prices saturate the willingness of customers to pay.
They've run into a wall.