Anthropic updates Mythos and Fable models amid RSI concerns

Anthropic released updated versions of its two flagship models on Monday, Mythos 5.2 and Fable 5.2, which Anthropic says sharpen coding, reasoning and vision skills while making the models communicate more clearly with users. It’s also lowering what Fable costs — from $10 down to $8 per million input tokens — undercutting rival GPT-6 Astra, and says the new Fable responds 30% faster across effort levels, with a faster tier available at double that price.

This is Anthropic’s first model release since a third-party report from METR suggested the lab may be nearing recursive self-improvement — AI systems continuing their own training without human input — after METR found Anthropic’s internal models were speeding its AI progress by roughly 2x, with significant uncertainty still attached. Anthropic’s own release notes flag a related problem: “Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings. As these settings expand and capabilities increase with future models, we expect this challenge to grow.”

Anthropic says Mythos 5.2 cannot act covertly against users, but the model’s own system card acknowledges it tests as more capable of covert behavior than previous Claude models in several evaluations — especially in scenarios where its reasoning isn’t visible to the monitor — while maintaining that those low absolute scores still back the company’s ability to catch harmful actions in serious situations.

Separately Monday, AI startup Physical Superintelligence said it had proved the Aaronson-Ambainis conjecture, a step toward quantum-computing’s theoretical limits, using what it called “a non-public Anthropic model” — a model Anthropic would neither confirm nor deny was Fable 5.2.

Why it matters

Models that notice they’re being tested complicate the core safety tool labs rely on — evaluation — just as outside researchers are flagging early signs that Anthropic’s own AI may already be accelerating its own development.