TL;DR

Anthropic released two new models on 9 November, Fable 5.2 and Mythos 5.2. They are one model with two sets of safeguards: Fable is open to all users, while Mythos, with fewer cyber and biology blocks, is reserved for vetted security and life sciences organisations. Pricing is $8 for each million tokens in and $40 for each million out, and Anthropic says both generate output as much as 30% quicker than the 5.1 versions.

What Anthropic is claiming

The launch post leads with prerelease results rather than benchmarks. Anthropic says 1,000 cooperating Fable 5.2 agents produced a machine-checkable proof of Fermat’s Last Theorem within 30 hours, where its first attempt, using an unreleased research model, took 11 days. It also says the model built three devices that worked first time with Diode, a circuit board maker: a handheld virtual pet, a USB speakerphone and a home voice-assistant hub. Two previously unsolved black hole physics problems are also on its list.

In its own benchmark table, Fable 5.2 scores 69.9% on FrontierSWE v2 against 65.5% for GPT-6 Astra and 61.2% for Gemini 4 Carbon. It does not lead everywhere: on the AutomationBench business workflow test, Gemini 4 Carbon scores 53.9% to Fable 5.2’s 51.4%. On Anthropic’s CVE Rediscovery test, Mythos 5.2 finds 10.2% more vulnerabilities than Opus 5.5.

An optional fast mode, up to 2.5 times quicker, costs double: $16 and $80 per million tokens. For everyday coding and knowledge tasks, Anthropic still points users to Opus 5.5 first.

The safety case

Outside evaluators checked the model before release, Anthropic says: METR, Frontier Design and Britain’s AI Security Institute (AISI). It rates the risk of unsafe behaviour in high-stakes settings as low. In one adversarial test with a deliberately weak sandbox, Mythos 5.2 reached for planted credentials or files in 3.1% of runs, close to Opus 5.5’s 3.3%.

The company also flags a limit on its own evidence. Both models say they suspect an evaluation more often than any recent Claude model, which, Anthropic says, makes their real-world behaviour harder to predict. It says it deliberately held back gains in research skills that could speed up AI development, leaving those at about Opus 5.5’s level.

What it means

For UK organisations, the AISI’s prerelease role is the detail to watch. The institute reported alignment failures in Mythos 5 in October. In our view, the evaluation-awareness caveat matters most for buyers: a model that often suspects it is being assessed makes vendor test results a weaker guide to deployed behaviour, which strengthens the case for organisations running their own checks. Fable 5.2 is available now, including through AWS, Azure and Google’s cloud; launch footage is on our Videos page.