Anthropic releases Claude Fable 5.2 and Mythos 5.2

The American AI company Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. They are the company's most powerful models, built for difficult technical jobs such as large programming projects and scientific calculations.

In fact, it is one model in two versions. Fable is available to everyone and has extra safeguards: it refuses dangerous tasks in computer security and biology. Mythos has fewer of these limits and can only be used by organisations that Anthropic has checked first.

Ahead of the competition, in its own tests

Anthropic compares Fable 5.2 in a table with GPT-6 Astra from OpenAI and Gemini 4 Carbon from Google. According to the tech site OfficeChai, Fable beats GPT-6 Astra on every test that includes a score for that model. Fable loses to Gemini 4 Carbon once: on a test of business workflows, Gemini reaches 53.9 per cent and Fable 51.4 per cent.

The biggest jump is on a test with real assignments of the kind freelancers get, the Remote Labor Index. There, according to Anthropic, Fable 5.2 successfully completed 41.3 per cent of the jobs, compared with 17.9 per cent for the previous version. On a programming test, Fable 5.2 performs about as well as Opus 5.5 at its most powerful setting, for about half the cost per task.

All the figures come from Anthropic itself. The models did not use the same settings in every comparison, and some scores for competitors are missing. Independent researchers now need to repeat the tests.

A famous mathematical proof in 30 hours

One striking result concerns Fermat's Last Theorem, a mathematical puzzle that was only solved in 1994. According to Anthropic, a team of 1,000 Fable 5.2 programs working together translated the complete proof into Lean in 30 hours. Lean is a language in which a computer can check every step of a proof. With the same set-up, Opus 5.5 reached 26 per cent after 18 hours and then made hardly any further progress.

Anthropic reports more. During tests before the launch, the company says, the models designed three working electronic devices, including a speakerphone for the computer, and solved two open questions about the physics of black holes. Mythos 5.2 correctly predicted the shape of 28 of the 34 capsids it studied. Capsids are protein shells used in gene therapy to carry genetic material into cells. Fable 5.2 also controls robots better: in a simulation with two robot arms it scored 64.9 per cent, compared with 39.4 per cent for Opus 5.5.

Availability and price

Fable 5.2 can be used in the Claude apps and through the cloud services of Amazon, Google and Microsoft. Organisations working on computer security or biomedical research can apply for access to Mythos through two special programmes run by Anthropic.

Both models cost 8 dollars per million tokens of text going in and 40 dollars per million tokens coming out. A token is a piece of text of about half a word to a whole word. That is cheaper than the previous version. A faster mode costs twice as much. For ordinary everyday programming and office work, Anthropic still recommends Opus 5.5.

Before the launch, the British AI Security Institute and the research organisation METR were among those who tested the models. Anthropic itself rates the chance of seriously unwanted behaviour as low. The testers did not sign off on that assessment themselves.

What does this mean for the road to the singularity?

AI models can work on their own on one large job for longer and longer, and a thousand of these programs can share out a task that used to take days. A mathematical proof that a computer has checked step by step shows that this work can also be trusted. If independent tests confirm the figures, AI that does its own research and designs things comes another big step closer.

What is the technological singularity? Read the explanation →