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.
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. My own experience is that this is true-- it is difficult to call Opus strictly better than Sol, but Sol gets where it's going far more efficiently.
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...
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
Note: This doesn't really capture the reality of things like 3d design or electronics design, where the earlier models just can't DO any of it well. Maybe you can bandaid over some of the gaps with strong skills, etc. The newer models can just do things that were not possible at all. When you move outside of coding into the visual/spacial space, you'll feel the inflection point on those capabilities that are happening right now.
> 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
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.
Doesn’t seem like a “wall” to me.
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.1
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.2
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.
1 - The process of distillation is super effective for specific domains though.
2 - I know I will be called out for supposed AI mystery machine. What I mean is that Large models rely on latent reasoning and nuanced context, often lost during condensation. Distilling models to use fewer tokens or parameters makes predicting retained details difficult. Some papers suggest that model preferences introduced during training, whether intentionally or inadvertently, may be inherited by distilled and fine-tuned models.
We would need to change how we compute. New chips. New architectures. Or maybe we reached the end of physics. Photonic compute is the last frontier. I don’t know.
They are doing IPO now because it might become apparent that they reached a limit in the next year
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.
And I am simply ignoring the ridiculous comments around local model usage from certain CEOs.
Same is happening in LLMs. Sadly, the Golden Era is over.