Anthropic launches Claude Fable 5, first public Mythos-class model with cyber safeguards

Claude Fable 5 is Anthropic's first general-availability model built on the Mythos architecture, and the company frames it as a major-version step change rather than an incremental bump. Independent voices agreed: Andrej Karpathy called it "SOTA on everything by a margin" and "a major-version-bump-deserving step change," while Anthropic's Claude Code lead Boris Cherny said it was "the biggest step up I've felt in our models since Opus 4.5." The model is designed for long-running autonomous agent work, sustaining complex coding and knowledge tasks for days without intervention, and natively understanding diagrams, charts and tables nested in documents.

The defining feature is safety gating. Anthropic says Fable 5's raw cyber capabilities are dangerous enough that queries touching cyber, bio and chemistry domains are automatically routed to the weaker Opus 4.8 — a hard limit baked into the public release. The launch arrived a week after a Trump executive order on voluntary pre-release AI safety testing, and Anthropic paired it with a public call for a frontier-development pause and an AI-nonproliferation argument.

Competitively, Fable 5's pricing doubled output rates versus Opus 4.8 ($10/$50 vs $5/$25), and benchmarks against Gemini 3.1 Pro and GPT-5.5 are contested. LlamaIndex's Jerry Liu found Fable 5 crushing on reasoning-heavy benchmarks like SWE-Bench Pro and GDPval but only matching Gemini 3 Flash on document parsing at 10-15x the token cost.

The unrestricted Mythos 5 remains gated to existing Project Glasswing partners. Skeptics question whether capability-gating-by-safety is genuine protection or marketing, especially given Anthropic's ties to government safety-testing programs and the apparent inconsistency of shipping Fable 5 while calling for a pause.