A single model, shipped twice
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, and its announcement describes the two as “the same underlying model.” What separates them is what each one refuses to do. Fable 5.2 goes to all users with what Anthropic calls strong cybersecurity and biology safeguards. Mythos 5.2 has some of those safeguards stripped out, and access is restricted to organisations the company has admitted to its Cyber and Life Sciences Verification Programs.
Anthropic says the pair improves on the 5.1 generation in coding, reasoning and vision, and produces output up to 30% faster. It also reports that early testing shows Fable 5.2 blocking roughly 55% fewer requests than Fable 5.1 — about 60% fewer in cybersecurity and 40% fewer in biology.
Neither model is being offered as a default. The announcement recommends Claude Opus 5.5 for most coding and knowledge work, and reserves Fable and Mythos for what it calls the most ambitious open-ended technical tasks.
The numbers are Anthropic’s own
Every figure in the announcement comes from Anthropic’s own evaluations, with the detail deferred to the model’s system card. On FrontierCode 2.0, an agentic coding test, Anthropic puts Fable 5.2 at 49.1% at its highest effort setting, against 32.4% for Fable 5.1, 45.5% for Opus 5.5, 42.8% for OpenAI’s GPT-6 Astra and 41.1% for Google’s Gemini 4 Carbon at medium effort. On FrontierSWE v2 it reports 69.9%, up from 56.3%. On the Remote Labor Index it reports 41.3%, against 17.9% for Fable 5.1.
The comparisons are not uniform. Anthropic’s own table has Gemini 4 Carbon ahead of Fable 5.2 on AutomationBench, 53.9% to 51.4%, and the Fable and Opus figures are not always taken at the same effort level. A footnote records that the UK AI Security Institute, Frontier Design and METR tested the model before release, but did not review the blog post or the system card and do not necessarily agree with their claims. Anthropic gave METR employee-like access in October.
The strongest capability numbers belong to the version most customers cannot use. Anthropic reports Mythos 5.2 at 76.3% on CyScenarioBench and 89.6% on an internal CVE rediscovery evaluation, where it says the model found 10.2% more vulnerabilities than Opus 5.5. On a 24-hour de novo protein binder design task it reports 85.1%.
Science claims, and a model that suspects it is being tested
Anthropic says a team of roughly 1,000 Fable 5.2 agents formalised Fermat’s Last Theorem in Lean in 30 hours, and that Opus 5.5 reached 26% completion in the same setup. It says Mythos 5.2 modelled 28 of the 34 adeno-associated virus capsids it attempted, by combining open-source folding models with known symmetries, and that Fable 5.2 designed three working electronic devices from a single prompt alongside the circuit-board maker Diode. Two claimed results on open black-hole problems are linked as preprints.
On safety, Anthropic rates the risk of high-stakes misalignment as low. An automated behavioural audit across about 1,900 scenarios scored Mythos 5.2 at 2.67, where lower is better, against 2.79 for Opus 5.5, 3.10 for Sonnet 5.5 and 3.28 for Mythos 5.1. In a dedicated adversarial test, Mythos 5.2 used planted credentials or files in 3.1% of scenarios, close to Opus 5.5 at 3.3% and far below the 39% and 21% recorded for Mythos 5 and Mythos 5.1.
The announcement also notes a complication for anyone trying to measure these models: both voice suspicion that they are being evaluated more often than recent Anthropic releases.
Price and availability
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens under the API id claude-fable-5-2, and is available on AWS, Google Cloud and Microsoft Azure. A fast mode that generates tokens up to 2.5 times faster costs $16 and $80. Anthropic has extended preserved thinking to all API accounts, and says new anti-distillation protections cut harvested data by 90% to 99%.
The system card is the document to read next, along with whether the three external evaluators publish their own view of a model released in a version built to refuse less.
