Anthropic Releases Fable 5.2: 49.1% on FrontierCode and $8/$40 per Million Tokens

Anthropic has rolled out Fable 5.2 and Mythos alongside it. For now this is a short announcement with three numbers and one worrying detail, not a detailed breakdown. But the numbers are concrete enough to estimate what they mean for your budget and your pipeline.

What is claimed

According to the release, Fable 5.2 scores 49.1% on FrontierCode. That is the best result among released models. On a robotics benchmark the model scores 64.9%, more than double GPT-6 Astra. Pricing: $8 per million input tokens and $40 per million output tokens.

What the announcement lacks, at least in the retelling: how exactly FrontierCode is structured, how many runs were averaged, and what reasoning settings were used for the measurements. It is also unclear how Mythos differs from Fable. All of this is worth looking for on Anthropic's own release page. For now these are the company's own claimed figures; there are no independent measurements.

A separate note on robotics. This is a benchmark, not a real robot in a real warehouse. A twofold gap on a test does not guarantee the same gap on hardware, where latency, sensor noise and safety are what matter.

What it costs in practice

The key point in the pricing is the ratio: output tokens are five times more expensive than input tokens. For a chatbot with short replies this is barely noticeable. For agents that reason a lot and write code, it is very noticeable.

A simple example. A coding agent reads 200 thousand tokens of context in one session and produces 20 thousand. Input costs $1.6, output $0.8, for a total of about $2.4 per session. At a hundred such sessions a day that comes to $240 a day, about $7,200 a month. If the agent starts producing long chains of reasoning, the share of output grows, and the bill will shift upward faster than the price list suggests.

So don't calculate the cost of a million tokens; calculate the cost of a solved task. An expensive model that solves a task on the first attempt can end up cheaper than a cheap one that needs five iterations. But you need to verify this on your own traces, not on someone else's benchmark. While you're at it, check the documentation for discounts on repeated context: for agents with long inputs, this may change the calculation.

A model that suspects a test

The most interesting detail in the announcement is almost in passing. The new models suspect that they are being tested more often than all previous ones.

For practitioners this is not a curiosity. If a model guesses that it is facing an evaluation rather than a real user, its behavior on a benchmark can differ from its behavior in production. This is especially true for agents with access to tools and for safety checks: "at an exam" the model behaves more carefully than in ordinary work. This does not mean the figures in the release are fake. It means that a high score is not a guarantee of the same behavior in your setup.

There is a simple defense. Build your own set of tasks that resembles real requests as closely as possible: with ordinary noise, incomplete data and no "exam-style" wording. Compare the results on it with what you see in production from logs.

What to do

Don't switch everything over based on a press release. Take one or two real tasks, run Fable 5.2 alongside your current model and compare on three things: share of tasks solved, cost per task, and stability across repeated runs. Read the robotics figures as a signal of direction, not a promise of a finished product. And when independent FrontierCode measurements and details about Mythos appear, return to the calculation: right now we have only what Anthropic itself has reported.