Anthropic launches Claude Fable 5 and Claude Mythos 5

Anthropic released Claude Fable 5, described as its most capable widely released model, alongside Claude Mythos 5 — a model Anthropic said shares Fable 5’s underlying capability but omits Fable 5’s safety classifiers, which can decline requests outright. Both are the second generation of Anthropic’s “Mythos-class” tier, sitting above the Opus line, following Claude Mythos Preview’s limited April release through Project Glasswing, a US government collaboration on cybersecurity applications.

Mythos 5 was not made generally available. Access was limited to existing Glasswing partners and vetted cyber-defence users, with a separate trusted-access programme for biology researchers planned to offer Fable 5 with its chemistry and biology safeguards removed but cyber safeguards intact. Anthropic said Fable 5 achieved state-of-the-art results on “nearly all tested benchmarks” it reported, including the top score on Cognition’s FrontierCode evaluation, and cited Stripe’s account that the model “compressed months of engineering into days” on a codebase migration. For Mythos 5, Anthropic said researchers preferred its scientific hypotheses over prior models’ roughly 80% of the time in evaluation, and that it had conducted autonomous genomics research it claimed outperformed a recently published Science paper — claims that came from Anthropic’s own announcement rather than independent replication.

Both models launched at $10 per million input tokens and $50 per million output tokens via the API, with Fable 5 initially bundled into Pro, Max, Team and Enterprise subscription plans before shifting to usage-credit billing later in the month.

Three days after launch, both models were pulled offline worldwide under a Commerce Department export-control order — the subject of a separate entry — an early test of how the restricted-access model for Mythos-class systems would hold up once a jailbreak was reported against the technique it depended on.