Anthropic ships Claude Fable 5, the first Mythos-class model open to the public

A fourth tier above Opus arrives with safety-class fallbacks, a 30-day retention floor, and a restricted Mythos 5 twin for Project Glasswing.

Anthropic released Claude Fable 5 on Tuesday, the first publicly accessible model in its new Mythos tier, alongside a restricted Mythos 5 twin reserved for Project Glasswing cyberdefenders and a forthcoming cohort of biology researchers. The release confirms what the April Mythos Preview already implied: Anthropic now sells the same base model in two shapes, one open and classifier-gated, one closed and unfiltered, with the difference legible mostly at the policy layer.

Pricing on both lands at $10 per million input tokens and $50 per million output tokens, roughly double Opus 4.8. Fable 5 ships with a 1M-token context window and up to 128k output tokens per request, available through the Claude API, AWS Bedrock, Vertex AI, and Microsoft Foundry, with subscription access through Pro, Max, Team, and seat-based Enterprise plans arriving June 22 pending capacity.

The numbers are loud. Per The Decoder, Fable 5 scored 80.3% on SWE-Bench Pro against Opus 4.8’s 69.2% and GPT 5.5’s 58.6%, and posted 29.3% on Cognition’s FrontierCode versus Opus’s 13.4%. Stripe told Anthropic the model finished a Ruby migration across a 50-million-line codebase in a day.

What’s structurally new is the safety architecture. Fable 5 carries classifiers over cybersecurity, biology, chemistry, and distillation domains; Mythos 5 doesn’t. When a classifier fires, the Messages API returns HTTP 200 with stop_reason: "refusal", and a new fallbacks parameter routes the request server-side to Opus 4.8. Anthropic says the trigger rate sits under 5% of sessions, tuned conservatively to ship the model quickly. A 1,000-hour external bug bounty against the classifiers, TechCrunch reports, surfaced no universal jailbreaks.

The tradeoff is retention. Both models override prior zero-retention agreements with a mandatory 30-day floor, which Anthropic says will be used only to defend against novel attacks and reduce false positives, not for training. It’s the same bargain regulators floated after the 2023 Executive Order on AI: capability access in exchange for telemetry. Anthropic is the first frontier lab to write it into the price sheet.

Anthropic ships Claude Fable 5, the first Mythos-class model open to the public

Sources