Anthropic introduces Claude Fable 5.2 and Mythos 5.2: cheaper API, robot control, and formal proofs

Claude Fable 5.2: more powerful, faster, and cheaper than Opus 5.5

On FrontierCode 2.0, Claude Fable 5.2 with High effort matched the score level of Opus 5.5 with Max effort, while the cost per task was almost twice lower.

The API price dropped from $10/$50 to $8/$40.

Robotics and physical control

The model is noticeably better at controlling physical robots, including tasks it was not specifically trained for. In demos, robots under its control assemble a cube, unscrew a bottle cap, and draw a robot. On the RoboDojo benchmark, the model scored 64.9%, compared with 29.8% for GPT-6 Astra.

Mythos 5.2 and biomedicine

Mythos 5.2 proposed a new method for modeling capsids — protein shells used for delivery in gene therapy.

Formalizing mathematics

A team of 1,000 Claude Fable 5.2 agents fully formalized the proof of Fermat’s theorem in Lean in 30 hours with almost no human involvement. The first such formalization at Anthropic took 11 days, required a special environment, and relied on researcher hints. The new run used three times fewer tokens, and the proof became seven times shorter.

Building an OS from scratch

In 8 days, the model wrote a working operating system from scratch that boots on a laptop. It includes a kernel, disk, keyboard, and network drivers, file storage, a window interface, a login screen, an installer, and embedded Claude.

Link: Claude Fable and Mythos 5.2.