San Francisco, 10 November. Anthropic released the models Claude Fable 5.2 and Claude Mythos 5.2 on 9 November, which, it says, are stronger at programming, reasoning and working with images, produce output up to 30% faster than their predecessors and cost less, the company said.
Fable and Mythos, Anthropic explains, are the same underlying model. Fable has strict safeguards in cybersecurity and biology, so it was opened to all users, while Mythos without those restrictions goes to organisations that have passed vetting under its cybersecurity and life-sciences verification programmes. Fable 5.2 is available in Claude, Claude Code and through the API, including on Amazon Web Services, Google Cloud and Microsoft Azure; for developers the model identifier is claude-fable-5-2. The price is 8 dollars per million input tokens and 40 dollars per million output tokens; a fast mode that speeds up output by up to 2.5 times costs twice as much.
The company ran the comparisons with rivals’ models itself, and they show both strengths and weak spots. FrontierCode 2.0 tests whether maintainers of open-source projects would accept proposed code changes. At its highest effort level Fable 5.2 scored 49.1% against 42.8% for OpenAI’s GPT-6 Astra and 41.1% for Google’s Gemini 4 Carbon (the latter tested at medium effort). In RoboDojo Sim, a simulation of controlling two-armed robots, the result was 64.9% against 29.8% for GPT-6 Astra. But on AutomationBench, a test of business workflows, Gemini 4 Carbon leads: 53.9% to 51.4%.
In pre-release testing, Anthropic says, Fable 5.2 solved two open problems in theoretical physics about black holes and, with about a thousand agents, formalised the proof of Fermat’s Last Theorem in 30 hours; the first formalisation, done by a research version of the model, took 11 days. The company’s announcement offers no independent verification of these claims.
Before release the model was tested by outside organisations: the UK AI Security Institute, Frontier Design and METR. Anthropic rates the risk of dangerous behaviour in critical situations as low, but admits a difficulty: both models voice suspicion that they are being evaluated more often than earlier ones, which makes it harder to know how they will behave in real conditions. In a sandbox-escape test Mythos 5.2 used data left outside the sandbox in 3.1% of scenarios, while Mythos 5 and 5.1 did so in 39% and 21%. The company also limited, in training, the growth of narrow skills that most directly speed up AI development itself: in those areas the models are at the level of Opus 5.5.
That last point ties into a METR report published on 5 November. The independent non-profit estimated that the capabilities of Anthropic’s models on well-specified tasks are growing about twice as fast as over the previous 18 months, Reuters reported. The estimate rests on Anthropic’s own internal index, AECI: an acceleration of 1.9 times with a confidence interval from 1.5 to 2.8, OfficeChai reported. METR considers the main cause to be the use of AI itself in development rather than growth in computing power. Some information in the report is withheld, so the calculation cannot be checked from outside.
What it means: Anthropic cut the price of its flagship line and released it four days after a METR report on the possible speed-up of AI development by AI itself. All the main comparisons with OpenAI and Google were run by the company itself; there are no independent measurements of the new models yet.
Who benefits: Developers and companies that pay for access: Fable 5.2, TechCrunch notes, is now cheaper than GPT-6 Astra. Vetted organisations: only they get Mythos 5.2 without restrictions in cybersecurity and biology. Anthropic itself: the release comes before a share offering that, according to the newspaper, is set for next week.
Background: Fable and Mythos are Claude’s senior lines above Opus. The first versions of Fable 5 and Mythos 5 came out on 9 June 2026; on 12 June Anthropic suspended access to them to comply with US Department of Commerce export controls, which the department lifted on 30 June, and access was restored on 1 July, the company said.
