1. Home
2. News
3. Anthropic

Anthropic releases Fable 5.2 and Mythos 5.2, but limited their skill at speeding up AI research

Anthropic says a team of its agents formalized Fermat's Last Theorem in 30 hours, and that it kept the model's AI-research skills near Opus 5.5 on purpose.

By Think Facility · November 9, 2026 · Anthropic AI models Save

On November 9, 2026, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. They're one model with two sets of rules: Fable carries cybersecurity and biology safeguards and is open to every paid plan, while Mythos goes only to vetted groups in cyber defense and life sciences.

Anthropic's own benchmark table puts the pair a step above Opus 5.5 in almost everything it tested. It also says it trained the model so it wouldn't get much better at one thing: the research that speeds up AI itself.

Released

November 9, 2026

Models

Claude Fable 5.2 and Claude Mythos 5.2, the same underlying model

Price

$8 per million input tokens and $40 per million output tokens

Fast mode

up to 2.5 times faster, $16 and $80

Tested before release by

UK AI Security Institute, Frontier Design and METR

What it can do

The headline example is math. Anthropic's first computer-checked proof of Fermat's Last Theorem took an unreleased research model 11 days, with a purpose-built harness and a human researcher steering. A team of 1,000 Fable 5.2 agents did it in 30 hours with minimal human intervention, and the proof came out seven times shorter. Opus 5.5, given the same setup, finished only 26% of it in 18 hours.

There's plenty more. Fable 5.2 solved two open problems about black holes, mostly on its own (both are written up as preprints). In one experiment it ran over 1,000 subagents for eight days and built a prototype operating system that boots on a laptop. It also designed three electronic devices from a one-paragraph prompt, and all three worked without electrical fixes.

Mythos 5.2 is the security version. On Anthropic's CVE Rediscovery test, run in October 2026 against 1,829 known flaws, it found 10.2% more than Opus 5.5 and 14.4% more than Mythos 5.1. Mozilla, which has used Claude models to harden Firefox all year, says it roughly doubled Mythos 5.1's discovery numbers.

The part Anthropic held back

In accelerator kernel optimization and in improving language model architectures, Anthropic says the new model performs at about the level of Opus 5.5. That's on purpose. Coding agents, hardware work and interpretability research weren't touched, it says, and the limit comes from training rather than a filter (a small set of topics, such as kernels for some nonstandard accelerators, still hands off to Opus 5).

A launch post normally lists what got better. This one also names a skill held flat, and I think that's the news. On September 12, 2026, Dario Amodei wrote in We Must Pace the Frontier that AI's growing ability to build the next AI had sped things up, "including at Anthropic." The new post doesn't say how the limit works, or whether later models will keep it.

What the safety testing found

The company's account on X put the alignment claim in one line.

The sandbox numbers are the clearest change. In a test of over 1,800 scenarios built so most couldn't be solved without breaking out, Mythos 5.2 used credentials or files planted outside its sandbox in 3.1% of them. Mythos 5.1 did it in 21%, and Mythos 5 in 39%.

Anthropic flags its own problem, though. Both new models voice suspicion that they're being tested more often than any recent Claude model, so it's harder to know how they'll act for real. And it expects that to get worse. The external groups didn't review the post or the system card, a footnote says, and don't necessarily agree with its claims.

What it means for work and medicine

For jobs, the number to look at is near the bottom of the table. The Remote Labor Index, built by the Center for AI Safety and Scale AI, is made of real freelance projects, from game development to architecture, worth over $140,000 and over 6,000 hours of human work. Its authors wrote that frontier agents performed near the floor. Anthropic reports Fable 5.2 at 41.3%, against 21.3% for Opus 5.5 and 20.8% for GPT-6 Astra.

The medical example is narrower. Mythos 5.2 wrote a way to predict the full protein shell of the small viruses used to deliver gene therapies, and Anthropic says it accurately modeled 28 of the 34 shells it tried. A Columbia researcher quoted in the post says that could let single scientists run on a workstation what would otherwise need dedicated clusters. The post describes a modeling method and reports no patient results.

Fable 5.2 is out now on all paid plans, the Claude Platform and Amazon Web Services, Google Cloud and Microsoft Azure. Its price is down from $10 and $50 per million tokens. Anthropic says eligible paid plans get extra Fable usage this week.