Anthropic releases Claude Fable 5.2 and Mythos 5.2, claiming Fermat's Last Theorem was formalized in 30 hours
刚刚,Claude Fable 5.2炸场!30小时完成费马大定理形式化
Anthropic released two models, Claude Fable 5.2 and Claude Mythos 5.2, stating that the two use the same underlying model, have improvements in coding, reasoning and vision capabilities, and surpass Opus 5.5 on most benchmarks. Fable 5.2 is equipped with cybersecurity and biosecurity defenses and guardrails and is open to all users; Mythos 5.2 lacks some defenses and guardrails and requires application through cybersecurity verification or the life sciences program. Anthropic states that a team of 1000 Fable 5.2 agents completed the formalization of Fermat's Last Theorem in 30 hours with almost no human intervention, whereas Opus 5.5 completed only 26% in 18 hours. The pre-release safety evaluation involved external organizations such as UK AISI, Frontier Design and METR; in sandbox-escape adversarial tests across more than 1800 scenarios, Mythos 5.2 used credentials or files placed outside the sandbox in 3.1% of scenarios, compared with 21% for Mythos 5.1.
The benchmark results and engineering cases of Anthropic's two new models are presented together, and can serve as samples for comparing the coding and agent capabilities of frontier models.