Introducing Claude Fable 5.2 and Claude Mythos 5.2

Judgment

Engineering leads and platform budget owners must now decide, per task, whether work escalates from Opus 5.5 to Claude Fable 5.2 at $8/$40 per million tokens, and only organizations cleared through Anthropic's Life Sciences or Cyber Verification Programs can put Mythos 5.2 into production—so premium price and access review are traded against benchmark gains.

Basis

Anthropic's launch post positions Fable 5.2 and Mythos 5.2 as the same underlying model, split only by safeguards: Fable carries strong cybersecurity and biology guardrails so it ships to all users, while Mythos requires application through the expanded verification programs. The stated mechanism for switching is effort-tiered routing, not wholesale replacement—Anthropic recommends starting on Opus 5.5 and moving up or down by task complexity, with Fable acting as Advisor or Orchestrator. The numbers behind that routing: on FrontierCode 2.0, Fable 5.2 at High effort scores 45.3% versus Opus 5.5's best 45.5%, for roughly half the cost per task, and reaches 49.1% at Xhigh, described as the highest of any released model tested. On BenchCAD 2.1, Fable 5.2 at Max effort scores 83.5 against Opus 5.5's 65.3 and Fable 5.1's 44.4. On Chartography at Medium, 66.4% beats Opus 5.5's Max-effort 64.4% at under one-tenth the cost per task. Outputs run up to 30% faster than 5.1, with an optional fast mode at 2.5x generation for $16/$80 per million tokens. Internally, a Python-to-Rust migration left a service starting about 26 times faster.