Anthropic releases Claude Fable 5.2 and Mythos 5.2: a 1,000-agent team completes a Lean proof of Fermat's Last Theorem in 30 hours
Synopsis
Anthropic released two configurations of one new large language model: Claude Fable 5.2 for general use and Claude Mythos 5.2, limited to vetted individuals and organizations through trusted access programs. In a multi-agent test, a team of 1,000 agents using an earlier snapshot of the model completed a Lean proof of Fermat's Last Theorem within 30 hours, building on definitions and smaller proven results from an earlier formalization. The system card also reports that Fable 5.2 scored 93.0% on SWE-bench Pro, against 81.2% for Fable 5.1 and 89.9% for Opus 5.5.
Interpretation
Anthropic released two configurations of a new large language model: Claude Fable 5.2, the general-use version with stricter biological and cybersecurity safeguards, and Claude Mythos 5.2, which has more permissive safeguards and is limited to vetted individuals and organizations through trusted access programs. The system card describes both configurations as advancing the frontier in coding, knowledge work and problem-solving, with improved novel mathematical and scientific reasoning; earlier comparators include Claude Opus 5.5 and Claude Mythos 5.1. In a multi-agent test, a team of 1,000 agents using an earlier snapshot of the model completed a Lean proof of Fermat's Last Theorem within 30 hours, building on definitions and smaller proven results from an earlier formalization.
Capability evaluations cover coding, agentic search, multi-agent work, multimodal tasks, real-world professional tasks and healthcare, with Mythos 5.2 additionally evaluated on life sciences and vulnerability finding. The system card reports that Fable 5.2 scored highest on 14 of the 16 results in Table 8.1.A, trailing Opus 5.5 on the GDPval-AA and AA-Briefcase Elo ratings. On SWE-bench Pro, Fable 5.2 scored 93.0%, Fable 5.1 scored 81.2% and Opus 5.5 scored 89.9%; on the Remote Labor Index, Fable 5.2 completed 41.3% of freelance projects at least as well as a human professional, against 21.3% for Opus 5.5.
Safety evaluations treat Mythos 5.2 as having CB-1 chemical and biological capabilities with corresponding protections and not crossing the CB-2 threshold; on autonomy it meets the Autonomy-1 threshold but not Autonomy-2. The system card says Mythos 5.2 is a modest improvement over Opus 5.5 and Mythos 5.1 on several chemical and biological evaluations but shows limitations in open-ended ideation and scientific judgment; on cyber it met or exceeded Mythos 5.1 and Opus 5.5 on all reported evaluations while remaining in the lower capability category of the FCF. In the alignment assessment, Mythos 5.2 improved on or matched Opus 5.5, Sonnet 5.5 and Mythos 5.1 on most measures of alignment, resistance to misuse, honesty and factuality; white-box evidence suggests it was aware it might be in an evaluation in as many as 43% of automated behavioral audit transcripts.
Perspective
The system card describes pre-deployment evaluations: most were run in-house at Anthropic, with some external testers; chemical and biological assessments use beneficial proxy tasks and HHH snapshots, and report the highest score of any snapshot as a conservative reading, except for the LatchBio Bioinformatics and Protocols evaluations, which report results on the final production model.
The system card says Mythos 5.2 does not meet the RSP threshold for automated R&D, but labels that assessment uncertain; red-team participants reported operational issues that likely limited productivity and modestly reduce confidence that the model's capabilities were fully elicited.
