Claude Fable 5

Anthropic launched Claude Fable 5 on 9 June 2026 as the general-use half of a pair: the same underlying model as Claude Mythos 5, wrapped in deliberately conservative classifieurs de sécurité. On certain topics the request is answered by Opus 4.8 instead; Anthropic reports the guardrails firing in under 5% of sessions.

The most detailed evidence comes from Jarred Sumner, who used a pre-release Fable 5 to rewrite Bun from Zig to Rust in 11 days (3 to 14 May 2026). The method mattered as much as the model: a mechanical port validated by an existing TypeScript test suite, roughly 50 dynamic workflows in Claude Code, and adversarial review by a second Claude that saw only the diff. Result: 6,502 commits, a diff of +1,009,272 lines, and about $165,000 in tokens.

Price is the other reason the model keeps coming back. At $10 and $50 per million input and output tokens, Fable 5 costs twice Opus 4.8, which Didier Girard uses to argue that token spend becomes a variable OpEx line in 2027 budgets. Artificial Analysis measured 1783 Elo on GDPval-AA, ahead of Opus 4.8 at 1615 and GLM-5.2 at 1524.

Thariq Shihipar of the Claude Code team put the constraint elsewhere: Fable 5 is the first model where output quality is capped by the human's ability to clarify unknowns. Claude Fable 5.1, released 1 September 2026, scored 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5, which now serves mainly as the baseline.