Claude Fable 5

Anthropic released Claude Fable 5 on 9 June 2026 alongside Claude Mythos 5: the same underlying model, one made safe for general use, the other with guardrails lifted for a small group of cyberdefenders and infrastructure providers. Pricing is $10 per million input tokens and $50 per million output, twice Opus 4.8. The classifieurs de sécurité on Fable 5 are deliberately conservative, and when they fire (under 5% of sessions on average) the request is answered by Opus 4.8 instead.

Two accounts explain why the model keeps coming back. Jarred Sumner used a pre-release Fable 5 to port Bun from Zig to Rust in 11 days, a mechanical port validated by the existing TypeScript test suite (60,624 tests) with a second Claude in a separate context reviewing only the diff. Cost: roughly $165,000 in tokens, against three engineers for a year. Thariq Shihipar of the Claude Code team called Fable 5 the first model where the quality of the work is capped by his own ability to clarify its unknowns, which moves the bottleneck from the model to the person instructing it.

The benchmark picture is split. Fable 5 leads GDPval-AA at 1783 Elo, ahead of Opus 4.8 (1615) and GLM-5.2 (1524), and takes the best frontier score on FrontierCode. GPT-5.6 Sol beats it on Terminal-Bench 2.1 at about a third less per task, while Claude holds roughly 15 points on SWE-Bench Pro. Didier Girard reads the $10/$50 tier as a rising ceiling in a market whose floor keeps dropping.