Fable 5.2: Anthropic limits kernel optimization, Gemini 4 Hydrogen arrives

On November 9, 2026, Anthropic released Claude Fable 5.2 and Mythos 5.2 while limiting accelerator kernel optimization; on the 12th, Google released Gemini 4 Hydrogen. All the numbers are the labs' own, and on November 11 Anthropic changed its sentence on the proof of Fermat's Last Theorem to say what it had given the model.

In short

• Claude Fable 5.2 and Mythos 5.2 (November 9): same model, with safeguards for Fable, verified access for Mythos; 8 and 40 dollars per million tokens; 49.1 percent on FrontierCode 2.0 at maximum effort against 42.8 percent for GPT-6 Astra, according to Anthropic.
• A safeguard few people discuss: the system card lists kernel development on certain accelerators among the capabilities blocked on Fable 5.2 (fallback to Opus 5), while training limits kernel optimization to about the level of Opus 5.5.
• Claude Code's /advisor command lets Haiku, Sonnet or Opus consult Fable 5.2 on hard decisions. Anthropic's published savings (April 9, 2026) used Opus 4.6 as the advisor, not Fable 5.2.
• Fermat: "formalized in 30 hours by 1,000 agents". On November 11, Anthropic corrects the announcement: the agents had been given 60 statements, a sketch of the classical route and 47 lemmas already proved in a first formalization.
• Gemini 4 Hydrogen (November 12): 1.50 and 7.50 dollars per million tokens over a 1 million token context; Artificial Analysis index of 51 against 50 for Argon, according to Google.
• Qwen 4 has existed since November 11 as a paid preview, with no published benchmark, weights or price, retired on November 25.

← All the watch