Anthropic on Tuesday releases Claude Fable 5, a public, “Mythos-class” AI model that earlier Mythos previews were not broadly available for. The company positions Fable 5 as its most capable Mythos-family model for wider use, citing strong performance on tasks including software engineering, knowledge work, and vision. Anthropic and independent testing cite benchmark gains versus earlier models; one analysis from Hex says Fable 5 is the first model it tested to score above 90% on a long-running complex analytics benchmark, and it reports a substantial score increase on SWE-Bench-Pro compared with prior releases.
To limit misuse, Anthropic adds safeguards that prevent the model from answering certain high-risk topics such as cybersecurity and biology. When requests fall into those categories, Anthropic says Fable 5 routes them to Claude Opus 4.8 or blocks the response, and it says such routing occurs in a small fraction of sessions. Anthropic also says it tested the model against attempts to bypass restrictions and that these attempts were not successful.
Anthropic also updates access for existing Mythos preview users, including a new “Claude Mythos 5” variant with safeguards lifted in some areas. The release follows recent public calls from AI executives, including Anthropic, for coordinated efforts to manage frontier model risks and pace.
