Anthropic launches Claude Fable 5.2 and Mythos 5.2 for unattended technical work

Anthropic releases two new flagship models

Anthropic has launched Claude Fable 5.2 and Claude Mythos 5.2, a pair of flagship models aimed at hard technical work. According to The Verge, the company says the new models improve on their 5.1 predecessors — released in September — in coding, reasoning and vision, and generate output faster.

The rollout is split. Claude Fable 5.2 is generally available now, while Claude Mythos 5.2 is limited at first to organizations Anthropic has vetted, The Verge reports.

Designed to run without supervision

Fable 5.2 is pitched at open-ended technical work. According to Vercel's changelog, the model can sustain lengthy, loosely specified tasks without a human overseeing every step, and it can take on two distinct roles: an orchestrator that delegates simpler work to smaller models, and an advisor that other Claude models consult on difficult problems. The Verge describes similar strengths, saying the model can coordinate teams of agents and carry long-running software projects through to completion.

Anthropic's own prerelease testing claims, relayed by The Verge, go well beyond typical benchmark talk. The company says the models helped its engineers finish software migrations it describes as previously impractical, designed three working electronic devices from scratch, solved two open problems in black hole physics, advanced robotics evaluations, and produced a formal proof of Fermat's Last Theorem within 30 hours. These are vendor claims and have not been independently verified.

Fable 5.2 arrives on Vercel's AI Gateway

Vercel announced the same day that Claude Fable 5.2 is available through AI Gateway under the model name anthropic/claude-fable-5.2. Developers can call it via the AI SDK, the Chat Completions, Responses and Anthropic Messages APIs, or through coding agents such as Claude Code, Codex and Cursor connected to the gateway using Vercel's CLI setup.

The integration exposes an effort parameter that ranges from low to max, with high set as the API default. Anthropic has enabled cybersecurity and biology safety classifiers on the model, and Vercel advises configuring fallback models — through a models array in the gateway's provider options — so that a request rejected by the cybersecurity classifier is retried against the next model in the list rather than failing outright.

There is a significant data policy caveat: Fable 5.2 does not support Zero Data Retention. Prompts and completions are stored for 30 days, although Anthropic says they are not used for training. Teams with ZDR enabled will receive an error instead of a response. Vercel notes that AI Gateway pricing mirrors Anthropic's own, with no markup and no platform fee on inference, including on bring-your-own-key requests.

Why it matters

The release signals where frontier model development is heading: not just better answers to single prompts, but models built to work unattended for hours on messy, ill-defined problems while coordinating other models along the way. Anthropic's prerelease claims — particularly the formalized proof of Fermat's Last Theorem and the black hole physics results — are the kind of headline accomplishments meant to argue that AI can contribute to serious research, and they will invite scrutiny from outside reviewers.

The operational details matter for adopters too. Restricting Mythos 5.2 to vetted organizations suggests Anthropic is being selective about who gets its most capable model, and the mandatory 30-day retention on Fable 5.2 is a real constraint for regulated teams that normally insist on zero-retention terms. For developers, the Vercel AI Gateway availability makes the model reachable through existing tools and APIs — provided their data policies allow it.