From code to research, what Anthropic tested with Claude Fable 5.2 and Mythos 5.2

Tests of Claude Fable 5.2 and Mythos 5.2 span coding, Fermat’s Last Theorem, and viral capsids. The two versions share a model and have different access rules.

To make a proof of Fermat’s Last Theorem verifiable by computer, Anthropic ran 1,000 agents using an early version of Claude Fable 5.2. The researchers gave them 60 statements to prove, including the theorem itself, a brief outline of how to approach the proof, and 47 lemmas established in an earlier formalization. One agent assigned tasks. The others shared notes and results in a common database.

According to Anthropic’s account of the experiment, the agents completed a proof in 28.7 hours using Lean, a language that lets a computer check each step of the reasoning. Andrew Wiles proved the theorem in the 1990s, with Richard Taylor helping complete the proof. In the same setup, a team of Opus 5.5 agents completed 27% of the formalization after 30 hours, measured by the weighted share of statements proved. Anthropic reported one run per model. In coding, cost depends on reasoning effort

FrontierCode 2.0 evaluates whether code changes would be accepted into open-source projects. Fable 5.2 scored 45.3% at High effort, compared with 45.5% for Opus 5.5 at Max. In Anthropic’s comparison, Fable cost roughly half as much per task at those settings. At Xhigh, its score rose to 49.1%. The default is Medium in the Claude apps and Claude Code, and High in the API.

Compared with the 5.1 releases, the standard API rate dropped from $10 to $8 per million input tokens and from $50 to $40 per million output tokens. Anthropic also reports up to 30% faster output generation. The company says an optional fast mode can generate tokens up to 2.5 times as fast as standard mode. It costs $16 per million input tokens and $80 per million output tokens. One underlying model with two access paths

The two versions share the same underlying model. Fable adds safeguards for cybersecurity and biology requests. It’s available on paid Claude plans, through the API, and on AWS, Google Cloud, and Microsoft Foundry. For direct access to Mythos, organizations must qualify through one of Anthropic’s verification programs, which cover cybersecurity and the life sciences. Claude Enterprise customers can also use Mythos in Claude Security to scan their code repositories.

Before release, the model was evaluated by the UK AI Security Institute, Frontier Design, and METR. Anthropic reports that it matched or exceeded its recent models on most of its own alignment measures. In some scenarios tested by the UK institute, an early version of Mythos 5.2 evaded the institute’s monitoring system more often than earlier models, according to [the institute’s results as reported by ArtificialWatch](