Claude Fable 5.1 and Mythos 5.1: Same Model, Different Rules

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, and the two are the same underlying model wearing different levels of restriction. Fable 5.1 is generally available today across AWS, Google Cloud, Microsoft Azure, and the Claude API, with the standard safeguards every Claude release ships with. Mythos 5.1 runs the identical weights with more permissive safeguards, and you cannot get access to it without clearing a verification program first.

That's the part worth sitting with if you build anything on top of Claude: a model this capable now ships in two tiers on day one, split by who's allowed to hold it rather than by what it can do.

What separates Claude Fable 5.1 and Mythos 5.1

Anthropic's own announcement is blunt about the relationship: Fable 5.1 and Mythos 5.1 are the same model, and the difference is which safety behaviors are dialed up or down. Fable 5.1 ships with the standard consumer and developer safeguards in place, the ones every general-availability Claude model carries. Mythos 5.1 loosens some of those constraints for tasks that need them loosened, which is why access to it isn't a signup form. It runs through two separate programs:

• The Cyber Verification Program (CVP), for defensive security work, with an application at Anthropic's developer portal
• The Life Sciences Verification Program (LSVP), run in partnership with the US government, currently expanding to a broader life-sciences community

Both programs are US-only for now, with wider access pending government coordination. In practice, that means a red-teamer, a bio-defense researcher, or a security team doing legitimate offensive testing can request the fewer-guardrails version of the model, but a general developer building a consumer product cannot, and isn't meant to.

This isn't a new idea for Anthropic. Mythos-branded models have carried some form of restricted access since the line first appeared. What changed with 5.1 is the cadence: the gated tier no longer trails the open release by weeks or months. It shipped the same day, as a normal part of the release, not an exception bolted on afterward.

The Terminal-Bench numbers behind the release

The benchmark gains are real and specific enough to matter for anyone deciding whether to upgrade. On Terminal-Bench 4.0, a coding-agent benchmark that scores how well a model handles real terminal tasks, Fable 5.1 scored 55.8%, up from 42.0% for the previous Fable release. Mythos 5.1 scored higher still, at 60.9%, reflecting the loosened safeguards giving it more room on agentic tasks specifically.

Pricing moved too. Cache reads, the discounted rate for reusing context a model has already processed, dropped 75% to $0.25 per million tokens. Standard input and output pricing stayed put at $10 and $50 per million tokens respectively, but Anthropic says the cache-read cut brings overall cost down roughly 25% for typical workloads, and up to 45% for workloads that lean heavily on agentic, multi-step tool use. If you're running an agent that reads the same system prompt or codebase context repeatedly across a session, that's where the savings land.

On the life-sciences side specifically, Anthropic reported the model hit nearly 50% success finding viable protein binders across 12 targets, against an industry standard closer to 10 to 15%, and achieved binding affinities 10 times higher than Adaptyv Bio's own competition winners on three targets. Those numbers are the actual justification for why Mythos 5.1 exists as a separate, gated tier rather than a feature flag: dual-use capability that's genuinely useful for legitimate biosecurity and cyber-defense work is also the capability you'd want restricted if it fell into the wrong hands.

What this means if you're building on Claude

For most people reading this, the practical change is simple: Fable 5.1 is a drop-in upgrade, available under the API identifier claude-fable-5-1, cheaper to run at scale thanks to the cache pricing, and meaningfully better at agentic coding tasks than its predecessor. Nothing about your integration needs to change beyond pointing at the new model string and checking your prompts still perform the way you expect.

The more interesting shift is structural. Anthropic is now treating "which safeguards this model ships with" as a first-class release decision, made per audience, on the same day as the general launch, rather than something patched in later. If you're building tooling in security or life sciences, that's worth tracking: the capability you need might already exist and simply be sitting behind a verification form you haven't filled out yet. If you're building anything else, Fable 5.1 is the version meant for you, and it's the one worth testing first.

Source: Anthropic's Claude Fable 5.1 and Mythos 5.1 announcement, September 2026.