Anthropic says it limited AI-research skills while training Claude Fable 5.2 and Mythos 5.2

Anthropic · Company announcement

Read the source (opens in a new tab)
What happened
Anthropic's November 9, 2026 launch post for Claude Fable 5.2 and Claude Mythos 5.2 says it restrained some AI research capabilities during the model's training.
Key points
  • Anthropic says it "took steps to limit advances in the specific research capabilities that could most directly accelerate AI development."
  • In accelerator kernel optimization and improving large language model architectures, both models perform "at about the level of Opus 5.5," it says.
  • It says the changes did not affect agentic coding, hardware development or traditional AI and machine learning research, including interpretability, and that it does not alter these capabilities at inference time.
  • Both models also hand a small set of topics, such as kernel development on certain nonstandard accelerators, to Opus 5 through "transparent fallbacks."
Evidence
Company launch post stating a training decision. It gives no evaluation figures for it; the system card's safeguards section (1.5) describes the fallbacks.
Uncertain
The performance comparison with Opus 5.5 is Anthropic's own assessment and has not been independently tested.

See it in the latest record