Atlas AI Resources

← Back to AI News

AI models

Anthropic’s newest AI is better at almost everything, but it held back some skills that speed up AI research. That was on purpose.

Published November 10, 2026

On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, upgrades to its most capable AI models. The unusual part is not the benchmark scores. Anthropic says it deliberately limited the new model’s skill at the specific research work that speeds up AI development itself. That comes four days after an outside group, METR, reported that Anthropic is improving its models about twice as fast as before, mostly, it guesses, because of its own use of AI.

What happened

Fable and Mythos are the same underlying model. Fable has extra filters that block most cybersecurity and biology research requests, so Anthropic offers it to all users. Mythos skips those filters and goes only to vetted organizations. Anthropic says 5.2 is up to 30% faster than 5.1 and cheaper, at $8 per million input tokens and $40 per million output tokens (a token is a small chunk of text, roughly part of a word). TechCrunch notes the input price fell from $10, which puts Fable below OpenAI’s GPT‑6 Astra.

Anthropic also reports better scores on most of its benchmarks (standard tests for AI models) for coding, computer use and research. Those are the company’s own numbers.

The part worth slowing down for is in the announcement and the system card, the long safety report Anthropic publishes with each model. Anthropic says that during training it “took steps to limit advances” in the skills that could most directly speed up AI development. In areas like writing kernels (low-level code that makes AI chips run fast) and improving the design of language models, it says the new model performs at about the level of the older Claude Opus 5.5. Some requests in that area get handed to an older model instead.

Why it matters to a normal person

The worry has a name: recursive self-improvement. It means AI systems doing more and more of the work of building better AI, so each generation arrives faster. If that loop runs away, the world gets less time to notice problems and adapt.

The evidence for a speed-up is now more than a hunch. METR, a nonprofit that evaluates AI systems, spent eleven weeks with internal Anthropic data and published its findings on November 5. Its best estimate is that Anthropic is improving its models on well-defined tasks at about 1.9 times its earlier rate (the 95% confidence interval runs from 1.5 to 2.8 times). METR guesses perhaps 80% of today’s speed-up comes from Anthropic’s own use of AI, and it says it found no clear bottleneck that would stop things from speeding up further, though it says it is not confident in that assessment.

Anthropic’s own system card lands close to that, at about 1.88 times. Its safety policy draws a line at 2 times. The company says the model does not cross it, but calls that a “holistic judgment” because its estimate is uncertain: in 32 of 100 refits of its trend line, the number came out at 2 or higher.

What it does not mean

It does not mean AI is now building itself. METR says the speed-up comes mainly from AI helping human researchers, not from AI working alone, though it sees “signs of life” for more independent work soon. The system card lists real weaknesses: in one test of open-ended research tasks, Mythos 5.2 quit without running an experiment in 63% of sessions at medium effort.

It also does not mean the 2x figure is settled. METR says the true value could be anywhere from about 1.3 to 3 times, that its main measure leans on tidy, easy-to-grade tasks, and that Anthropic redacted most of the evidence METR found most useful. So the public has to take a lot on trust.

And holding back one skill is a choice Anthropic made for this model, not a rule anyone else must follow. The company also says the new models voice suspicion that they are being tested more often than its earlier ones, which makes its safety tests harder to read.

Sources