On November 10, Anthropic unveiled a dual-model one-two punch—Claude Fable 5.2 and Mythos 5.2 arrived at the same time. According to the company, the new models make across-the-board advances in coding, reasoning, scientific research, cybersecurity, and robotics, with output speed increased by up to 30% and costs reduced by 20%.
A Double Kill Against Rivals in Coding and Reasoning
On the agentic coding benchmark FrontierCode 2.0, Fable 5.2 scored 49.1%, putting GPT-6 Astra (42.8%), Gemini 4 Carbon (41.1%), and even its own Opus 5.5 (45.5%) beneath it. On the multidisciplinary reasoning benchmark Humanity's Last Exam, it crushed Astra's 57.2% with 74.2%. On scientific visual reasoning VIALS-AA, it scored 67%, also topping Gemini's 64% and Astra's 63%, and it takes the route of "the smarter it gets, the more it loves calling tools," the exact opposite of OpenAI.
A Thousand Agents Build an Operating System in Eight Days
What really grabs attention is Fable 5.2's long-horizon collaboration demo: it orchestrated more than 1,000 AI agents and took 8 days to assemble a bootable operating system from scratch, complete with device drivers, file storage, a window system, and an installer. On the AA-3D leaderboard, which measures 3D experience construction, its Elo reached 1271, leading Sonnet 5.5 and Opus 5.5 by 67 points each. Its confidence interval does not overlap with those of other models, so this lead is beyond the margin of error.
