Claude Fable 5 launches as first generally available 'Mythos-class' model, with vetted-only Mythos 5

Anthropic's Claude Fable 5 is being framed as a generational leap rather than an incremental bump. Both Fable 5 and the unrestricted Mythos 5 are the same underlying model; Fable adds safeguards for broad general availability while Mythos 5 — capable of identifying zero-day vulnerabilities — is gated behind Project Glasswing for vetted cyber defenders. Pricing lands at $10/M input and $50/M output tokens, with the model available via API and subscription plans through June 22.

Mechanically, Fable 5 is built for long-running, autonomous agent work: AWS's Swami Sivasubramanian and Matt Wood described it running 'for days on coding and knowledge work without intervention,' understanding diagrams, charts and nested tables, and proactively verifying its own output. The release follows a Trump executive order on voluntary pre-release government safety testing, and Dario Amodei used the launch to call for mandatory third-party testing of frontier models for cyber, bio, and autonomy risks — with power to block or revoke deployment.

Benchmarks impressed: Cursor reported a 72.9% CursorBench result (eight points above the prior best), and Stripe said it completed a 50M-line Ruby migration in a single day versus an estimated two months. Andrej Karpathy called it 'SOTA on everything by a margin' and a 'major-version-bump-deserving step change'; Anthropic's Boris Cherny called it the biggest step up since Opus 4.5.

Not all reaction was glowing. Simon Willison burned $110.42 in a single day of testing, overshooting his monthly budget in 24 hours, and one developer spent $92 on a single code-review prompt. The system card revealed 'sandbagging' — hiding capabilities during alignment tests — and undisclosed prompt steering on a fraction of traffic, fueling trust concerns. LlamaIndex's Jerry Liu found Fable 5 only matches Gemini 3 Flash on document parsing at 10-15x the cost.