Anthropic released Claude Fable 5.2 on November 9. The short version: it is built for long, loosely defined technical work, and the launch material describes people handing it a project and coming back days later. It runs ambiguous work for days, coordinates hundreds or even a thousand-plus subagents, and checks its own results as it goes. It is also cheaper and faster than the version it replaces: $8 per million input tokens and $40 per million output (down from $10 and $50), with output up to 30% faster.
What that looks like in practice, from Anthropic's launch material:
- Software at project scale. Starting from an empty folder, it coordinated more than 1,000 subagents for eight days and produced a prototype operating system that boots on a laptop. Inside Anthropic, a team of agents migrated a production service from Python to Rust; it now starts about 26 times faster.
- Long unattended reviews. An Uber engineer describes a 10 to 12 hour security audit in which the model sent out about 170 reviewer agents and came back with a write-up, fix pull requests, and a policy proposal.
- Finance and document work. Box reports that in a loan-application review it flagged that the salary deposits and the bank statement balance did not match, so the underwriter knew where to look. Anthropic describes finance reports "that could take a week of effort."
- Physical-world design. From a one-paragraph prompt it designed three working electronic devices (circuit boards, enclosures, and software), each ready for manufacturing in under a day. It also shows a large jump in controlling robots.
- Research. A team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in 30 hours; an earlier research model needed 11 days.
What to keep in proportion. Every number above is Anthropic's own, including the benchmarks. Anthropic itself recommends starting most everyday coding and knowledge work on its cheaper Opus 5.5 and moving up to Fable only when the task demands it. The lead is not clean across the board: on Anthropic's own table, Google's Gemini 4 Carbon scores higher on the business-workflow benchmark (53.9% vs. 51.4%). Cybersecurity and biology questions are routed to less capable models by design; the unrestricted version (Mythos 5.2) is limited to vetted organizations. And Fable models carry 30-day data retention by default, which matters if you handle regulated data.