Claude Fable 5.1 Arrives With a Gated Twin, Mythos 5.1
2 min readAnthropic shipped Claude Fable 5.1 on September 1, and it arrived with a sibling most people will never get to use. Claude Mythos 5.1 is the same model with looser safeguards, released only to vetted partners working in cybersecurity and the life sciences.
Two Models, One Set of Weights
Fable 5.1 is generally available now through the Anthropic API and on Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.1 goes only to registered participants in a trusted access program Anthropic calls Project Glasswing, where the safety classifiers and fallback restrictions that constrain Fable 5.1 have been removed.
The reasoning is that the same guardrails which stop a general purpose model from writing exploit code or working through pathogen biology also block the researchers whose actual job is exactly that. Rather than loosen the public model for everyone, Anthropic is loosening a private copy and controlling who holds it.
What Changed in Claude Fable 5.1
The headline gain is agentic science. Terminal-Bench-Science 0.1 more than doubled, from 24.7 percent on Fable 5 to 52.6 percent, and Anthropic says the model now leads Claude Opus 5 on every benchmark it published, reversing rows where Opus 5 previously won. Both models carry a one million token context window and can return up to 128,000 tokens in a single response.
Pricing is the quieter story. Input holds at $10 per million tokens and output at $50 per million, unchanged from Fable 5. Cache reads fell 75 percent, down to $0.25 per million. Anthropic estimates that lands at roughly 25 percent lower cost for typical workloads and up to 45 percent for heavily agentic work, since agents re-read the same context repeatedly. The company also says it addressed the verbose, jargon-heavy writing style users complained about in Fable 5.
Why It Matters
A cache-read discount is not glamorous, but it is the line item that decides whether long-running agents are affordable to operate. Cheaper re-reads make it rational to leave an agent working on one problem for hours rather than restarting it to keep the bill down.
The Mythos tier is the part worth watching. Gating capability by who you are, rather than by what the model refuses to do, is a real governance bet, and it holds only as long as the access control holds. That question got sharper this week: OpenAI’s report on the Hugging Face breach described what happened when highly capable models ran under reduced safeguards, per Axios.