Claude Mythos 5

On 9 June 2026, «Anthropic» launched «Claude Mythos 5» alongside «Claude Fable 5»: the same underlying model, with the guardrails lifted in certain domains. «Claude Fable 5» went to the public. «Claude Mythos 5» went to a small group of cyberdefenders and infrastructure providers, deployed initially through «Project Glasswing» in collaboration with the US government, as an upgrade to «Claude Mythos Preview». «Anthropic» credits it with the strongest cybersecurity capabilities of any model in the world. Both are priced at $10 per million input tokens and $50 per million output tokens, under half the price of «Claude Mythos Preview».

What «Anthropic» reports about «Claude Mythos 5» comes mostly from the life sciences. Its protein design experts accelerated their process by roughly 10x, and 9 of 14 protein targets produced strong candidates. «Anthropic» calls it the first model to generate novel and convincing scientific hypotheses, preferred around 80% of the time in blind comparison, and describes autonomous genomics research in which it trained a model 100x smaller that outperforms a recent Science publication. Automated alignment evaluation places its misaligned behaviour at a low level, similar to Opus 4.8.

The reason «Claude Mythos 5» keeps resurfacing is the access question rather than the benchmarks. A 7 August 2026 reading of Mistral AI's Shieldstral classifier uses the split as its foil: «Anthropic» keeps the guardrail inside the weights and decides itself who is allowed past it, while Mistral ships the guardrail as a separate, open, self-hostable component. Two models, one set of weights, and a vendor holding the key.