Anthropic releases Claude Fable 5.2 and Mythos 5.2, faster and cheaper than 5.1

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026 and calls them its most capable models so far. The two are one underlying model sold in two forms that differ in safeguards. Fable carries protections around cybersecurity and biology and is available on every platform. Fable 5.2 is on all platforms. Mythos has looser guardrails and goes only to vetted organisations through the company's Life Sciences and Cyber Verification Programs, which organisations can apply to join now. Anthropic says the models suit the most ambitious, open-ended technical work.

This is the second update to the Fable 5 line, which began in June, after Fable 5.1 in September, and the class itself began when Fable 5 made a Mythos-class model safe for general release; Claude Opus 5.5 arrived in July. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, about 20% less than the previous generation, and it generates output up to 30% faster. An optional fast mode runs up to 2.5 times quicker, at $16 and $80 per million tokens.

“Over 90% of the changes merged into Anthropic’s core codebase are now written by Claude (with human guidance).”

Anthropic, in its launch announcement · anthropic.com ↗

Anthropic ran the new model against Fable 5.1 on 14 benchmarks and beat it by more than 10% on most. It scored 74.6% on Terminal-Bench-Science 0.1, a 22% improvement, and 49.1% on FrontierCode 2.0 at the highest effort setting, the best result of any released model Anthropic tested. It also gained more than 13% on two coding benchmarks, FrontierCode 2.0 and FrontierSWE 2.0. At the High effort level it scored 45.3% on FrontierCode 2.0, against 45.5% for the best Opus 5.5 result, for about half the cost per task. On the hardware-engineering benchmark BenchCAD 2.1 it reached 83.5, against 65.3 for Opus 5.5 and 44.4 for Fable 5.1.

The launch leans on long agent runs. About 1,000 agents produced a formal, computer-verifiable proof of Fermat's Last Theorem in 30 hours, where an earlier unreleased model had needed 11 days, and the model answered two open questions in black-hole physics. One engineer had more than 1,000 subagents build a prototype operating system over roughly eight days; it boots on a laptop with network access and a Claude sidebar. A team of agents also moved a production service from Python to Rust, and the result starts about 26 times faster.

In biology, Mythos 5.2 accurately modelled 28 of 34 AAV capsids, a task Google DeepMind's AlphaFold2 cannot do because it handles single proteins, even though AlphaFold's creators shared the 2024 chemistry Nobel Prize. Fable 5.2's safeguards cause 55% fewer false blocks than 5.1, and prompts that are blocked on biology or cyber grounds are routed to Claude Opus 4.8 or Claude Opus 5. Anthropic says it believes high-stakes misalignment risk stays low and that its protections against distillation attacks are the strongest it has had. A system card with the full methodology is published, and Anthropic says the Rust-migrated service was a production one.

For everyday work, Anthropic suggests starting most coding and knowledge tasks on Opus 5.5 and moving up to Fable or Mythos for the most open-ended technical jobs, or down to Claude Sonnet 5.5 or Claude Haiku 5.5. Opus 5.5 can ask Fable or Mythos for advice when it is stuck, or Fable can direct smaller models. Eligible paid-plan subscribers get one to three Fable usage resets depending on tenure, and Fable counts for up to 50% of weekly limits on Pro and standard Team seats.