Claude Fable 5.2 and Mythos 5.2: what Anthropic's November 9 launch claims, and what is only Anthropic's word

Anthropic released two new models on Monday, November 9, 2026: Claude Fable 5.2, available to everyone, and Claude Mythos 5.2, available to vetted organizations through its Cyber and Life Sciences Verification Programs. According to Anthropic's launch post, they are the same underlying model. Fable carries cyber and biology safeguards. Mythos does not.

The headline claim is that this is Anthropic's most capable model for demanding technical work. The tables in the launch post support a narrower reading than that, and it is worth going through them.

Price, speed and where it runs

Anthropic lists standard pricing at $8 per million input tokens and $40 per million output tokens. A fast mode costs $16 and $80 and generates tokens up to 2.5 times faster. TechCrunch reports that Anthropic cut Fable's input price from $10, which puts it below its main competitor, GPT-6 Astra. Anthropic also says outputs arrive up to 30% faster than Fable 5.1 and Mythos 5.1.

The API model string is claude-fable-5-2, and Anthropic says the model is on Amazon Web Services, Google Cloud and Microsoft Azure. Mythos fast mode is restricted to organizations in the verification programs. Anthropic's own advice is to start with Claude Opus 5.5 for most work and move up to Fable or Mythos only for the most ambitious tasks.

What the benchmark table says

In Anthropic's table, Fable 5.2 scores 49.1% on FrontierCode 2.0, an agentic coding test, at Xhigh effort. Opus 5.5 scores 45.5%, GPT-6 Astra 42.8% and Gemini 4 Carbon 41.1% at Medium effort. On FrontierSWE v2 the figures are 69.9% for Fable 5.2, 65.5% for GPT-6 Astra and 61.2% for Gemini 4 Carbon.

The big jumps are against Anthropic's own previous model. On Remote Labor Index, Fable 5.2 scores 41.3% against 17.9% for Fable 5.1. On RoboDojo Sim, 64.9% against 27.6%. Those are large gains in a single point release, and they are exactly the sort of number that deserves an outside replication before anyone plans around it.

It is not a clean sweep either. On AutomationBench, a business workflow test, Gemini 4 Carbon scores 53.9% and Fable 5.2 scores 51.4%. Anthropic prints that row in its own post. On Anthropic's own evaluations it also says vulnerability identification is roughly on par with Opus 5.5, in a footnote.

Every figure above is a vendor number. Anthropic names the UK AI Security Institute, Frontier Design and METR as external testers, and then notes that those organizations did not review the blog post or the system card. That is a useful sentence for readers: the outside testers are involved, but nobody outside Anthropic has signed off on these tables.

The cyber and science claims

For Mythos 5.2, Anthropic reports 89.6% on CVE Rediscovery, an internal run of 1,829 vulnerabilities in October 2026, against 78.3% for Mythos 5.1 and 81.3% for Opus 5.5 without safeguards. On its CyScenarioBench the figure is 76.3%, against 61% for GPT-6 Cyber.

The science section is the most striking and the hardest to check. Anthropic says Fable 5.2 solved two open problems in theoretical physics tied to holographic duality, with preprints linked. It says 1,000 Fable 5.2 agents completed a Lean formalization of Fermat's Last Theorem in 30 hours, where Anthropic's first attempt with an unreleased research model took 11 days. It reports a prototype operating system built by coordinating more than 1,000 subagents over eight days.

These are demonstrations chosen by the vendor. They show what the model can do on its best days, not how often it does it. The physics preprints and the Lean proof can be checked by outsiders, which is the right kind of claim to make, and we will watch for that review.

The recursive self-improvement backdrop

TechCrunch notes this is Anthropic's first release since METR's assessment of AI research acceleration at the company. Anthropic says it has limited certain AI-development capabilities, specifically accelerator kernel optimization and LLM architecture work, to roughly Opus 5.5 level, while leaving agentic coding, hardware development and ML research untouched. Our earlier explainer on recursive self-improvement sets out why that distinction matters.

Anthropic also says that over 90% of the changes merged into its core codebase are now written by Claude with human guidance. That is a company statement about its own workflow, and METR's work on the same question is covered in our piece on METR's AI R&D acceleration assessment.

What customers say, and what they do not

Launch posts quote happy customers, and this one does. Vercel's Joe Haddad says, "The model is in my head. It understands the intent behind every prompt I've given it." Shopify's Tom Pritchard says it feels like working with another staff-level engineer. Harvey reports 91% on its own eval versus 86% for the next-best model. These are partners with early access and an interest in the relationship.

Access, safeguards and who gets what

Anthropic says Fable 5.2 blocks about 55% fewer queries than Fable 5.1 through its safeguards, about 60% fewer in cybersecurity and 40% fewer in biology. Blocked cyber queries fall back to Opus 4.8 and blocked biology queries to Opus 5. A private beta of Enterprise Frontier Safeguards, with Bridgewater, Salesforce and Stripe named as participants, works with zero data retention, though customers outside that arrangement have 30-day retention for safety monitoring. For consumers, Anthropic offers eligible paid plans one to three Fable resets and includes Fable for up to 50% of weekly usage limits on Pro plans.

Our take

Fable 5.2 looks like a strong release at a lower price, and for coding agents it is worth a trial. The claim of the best model for hard technical work rests on tables Anthropic wrote and ran. We would treat 49.1% on FrontierCode 2.0 as a hypothesis until independent leaderboards agree, and we would read the safety side of the system card as closely as the benchmarks. The launch itself says both models voice suspicion of being evaluated more often than any recent Anthropic model, which complicates every score above.