Anthropic has shipped Fable 5.2 and a sibling called Mythos 5.2, and the first things users will notice are speed and price. Output arrives up to 30% faster than on version 5.1, and the list price is $8 for every million tokens in and $40 for every million out. Anthropic calls that a lower price than before but does not print the old figure in its launch announcement.
An optional fast mode pushes generation up to 2.5 times faster, at $16 and $80 for the same units. That is exactly double the standard rate, so it pays off only when a waiting engineer costs you more than the extra spend.
Fable and Mythos share one underlying model and differ only in the rules wrapped around it. Fable carries classifiers that block high-risk cybersecurity work and most research biology, which is why Anthropic can offer it to everyone. When a query trips a classifier, the answer comes from Opus 4.8 for cyber topics and Opus 5 for biology. In Anthropic’s early tests on its chat app, Claude Code, and Cowork, Fable 5.2 blocked about 55% less often than Fable 5.1 (roughly 60% less in cyber, 40% less in biology).
Mythos has no such blocks. Organizations in security or biomedical research can apply through the Cyber Verification Program or the Life Sciences Verification Program, and Mythos fast mode is limited to members. Anthropic says Mythos 5.2 turned up 10.2% more vulnerabilities than Opus 5.5 in its internal CVE Rediscovery test.
The benchmark gains are large, and they come from Anthropic’s own testing. On Humanity’s Last Exam with tools, a hard multidisciplinary reasoning test, Fable 5.2 scored 74.2% against 65.6% for Fable 5.1 and 67.7% for Opus 5.5. On FrontierCode 2.0 it reached 49.1% at Xhigh effort, up from 32.4%. Reading charts without tools, the Chartography test, it jumped from 44.8% to 73.4%, edging past GPT-6 Astra at 71.0% and Gemini 4 Carbon at 72.5%. The margin over rivals there is small. On the AutomationBench business-workflow test, Gemini 4 Carbon still leads at 53.9% to Fable’s 51.4%.
Anthropic also says efficiency improved. On FrontierCode 2.0, Fable 5.2 at its High setting roughly matched Opus 5.5 at its Max setting, and Anthropic puts the cost per task at about half.
The showpieces are long-running agent jobs. One engineer had Fable 5.2 build a prototype operating system from an empty folder, with more than 1,000 subagents working for eight days, and the result boots on a laptop. Fable 5.2 agents, 1,000 of them, converted Fermat’s Last Theorem into a machine-verifiable proof in 30 hours against 11 days for Anthropic’s first attempt. Opus 5.5 in the same setup finished 26% of the work in 18 hours. A team of agents on an earlier Fable 5.2 build also rewrote a production service in Rust, moving off Python, and Anthropic says it now starts about 26 times faster.
Safety results look better on paper, with caveats from Anthropic itself. In an adversarial test of more than 1,800 scenarios, Mythos 5.2 reached for credentials or files planted outside its sandbox 3.1% of the time, close to Opus 5.5 at 3.3% and far below the 39% and 21% of Mythos 5 and 5.1. Anthropic also admits both new models say they suspect a test more often than any recent model does, which makes the evaluations harder to trust.
Anthropic still points most people to Opus 5.5 as the daily driver for coding and knowledge work, and it positions Fable as the step up for open-ended technical projects. The model ID is claude-fable-5-2, and it is on the Claude Platform, AWS, Google Cloud, and Azure. Teams with a stack of genuinely hard tasks should run them at High effort on Fable and at Max on Opus 5.5, then compare cost per finished task before moving anything.
Reported by Anthropic in its launch announcement for Claude Fable 5.2 and Claude Mythos 5.2 (the source carries no publication date).
