Claude Fable 5.2 accelerates code and targets the most demanding tasks

Anthropic launches Claude Fable 5.2 and Claude Mythos 5.2, two versions of the same model, with announced gains in code, reasoning, and vision. Availability extends to major clouds, with more open access for Fable and restricted for Mythos.

Anthropic unveiled Claude Fable 5.2 and Claude Mythos 5.2 on November 9, two versions of the same model designed for the most demanding technical tasks. Fable is open to everyone thanks to enhanced cybersecurity and biology safeguards, while Mythos remains reserved for verified organizations. The company also promises responses up to 30% faster than with version 5.1, with reduced pricing on the API. The message is clear: the model aims to work faster, but not for everyone, which at least prevents another slip-up in the family of large models.

Two versions, one engine, different doors

Fable 5.2 and Mythos 5.2 share the same technical base. In practice, Anthropic has separated uses by adding extra protections to Fable to allow broad access, while Mythos goes through dedicated verification programs for cybersecurity and life sciences. This organization matters because it says something about Anthropic’s strategy: opening the general model while maintaining a more controlled path for sensitive work.

Availability has been announced since November 9 on AWS, Google Cloud, Microsoft Azure, and the Claude Platform, under the identifier claude-fable-5-2. Eligible subscribers to paid plans also receive additional usage this week, with one to three resets depending on their seniority. Fable can represent up to 50% of the weekly limits of Pro plans and standard seats of the Team plan.

Regarding API pricing, both models are charged $8 per million input tokens and $40 per million output tokens. An optional fast mode, presented as capable of accelerating generation up to 2.5 times, raises these amounts to $16 and $80. Anthropic nevertheless recommends starting with Claude Opus 5.5 for most development and knowledge tasks.

The code improves, but benchmarks don’t all tell the same story

Anthropic highlights clear gains in programming. On FrontierCode 2.0, Fable 5.2 scores 49.1%, compared to 32.4% for Fable 5.1, 45.5% for Opus 5.5, and 42.8% for GPT-6 Astra. With a high effort level, the model would be on par with Opus 5.5 for about half the cost per task. On AutomationBench, which evaluates business workflows, it achieves 51.4%, behind Gemini 4 Carbon’s 53.9% [to be verified].

Tests published by Anthropic also signal real usage: over 90% of modifications integrated into its main source code are now reportedly written by Claude, under human supervision. In an internal experiment, Fable 5.2 coordinated more than 1,000 sub-agents for eight days to produce a prototype operating system capable of booting on a laptop. Again, the interest is not only speed. It is the ability to distribute a large project without everyone going off in their own direction.

The Fermat demonstration, a highly calculated showcase

To mark the occasion, Anthropic also highlighted a spectacular mathematical experiment: nearly 1,000 agents formalized in 30 hours, in Lean, the proof of Fermat’s Last Theorem, already demonstrated by Andrew Wiles in 1995. Formalizing here means making the proof verifiable line by line by a proof assistant, without implicit shortcuts. So this is not a new demonstration, but a large-scale coordination test.

Details matter. The agents worked without Internet, divided into teams of 16 or 17 with a captain responsible for breaking down the task. Anthropic indicates they had 60 statements to prove, a sentence summarizing the approach, the necessary definitions, 47 results already proven during a previous formalization, and the Mathlib library. The last lock, Langlands-Tunnell, was sent to about 750 assistants despite very low chances as judged by the team leader. In the end, the final theorem was completed after 28.7 hours.

This demonstration says less about Fermat than about how Anthropic wants to position Fable 5.2: as a model capable of managing hundreds of agents on long, structured, and verifiable tasks. It is also an implicit response to the competition. Google DeepMind claims a gold medal level at the 2025 International Mathematical Olympiad, and OpenAI asserts advances on Navier-Stokes and Yang-Mills. The mathematical field has become a showcase, not just an academic exercise.

Cybersecurity, biology and research under control

Mythos 5.2 is presented as Anthropic’s most powerful model for cybersecurity engineering. In an internal evaluation covering 1,829 CVEs, it reportedly identified 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. At Mozilla, Brian Grinstead, Distinguished Engineer, estimates that Mythos 5.2 has “about doubled” the vulnerability discovery metrics of Mythos 5.1.

Anthropic also claims that the risk of dangerous behaviors remains low in high-stakes situations, after evaluations conducted by the UK AI Security Institute, Frontier Design, and METR. Most published results, however, come from internal evaluations, which calls for cautious reading. At the same time, specific accesses exist via the Cyber Verification Program for cybersecurity and the Life Sciences Verification Program for life sciences researchers. Claude Enterprise clients can also use Claude Security, which relies on all the publisher’s models, including Mythos 5.2.

The methodological point remains: Anthropic presents these results as signs of maturity but also notes that its agents still give up too quickly on difficult problems and poorly assess their difficulty. In other words, they are better at breaking down the task than they are at knowing when to persist. This may be the real current limitation.

By the numbers

  • November 9 — date of presentation of Fable 5.2 and Mythos 5.2.
  • 30% — announced increase in response speed compared to version 5.1.
  • 49.1% — Fable 5.2’s score on FrontierCode 2.0, according to Anthropic.
  • 1,829 CVEs — base of Mythos 5.2’s cybersecurity evaluation.
  • 1,000 agents — workforce mobilized for the formalization of Fermat’s Last Theorem.

To read