Claude Fable 5.1 arrives in Microsoft Foundry alongside restricted Mythos 5.1

Anthropic released Claude Fable 5.1, a public preview model now listed in the Microsoft Foundry model catalog, alongside Claude Mythos 5.1, a restricted version that retains full cyber and bio capabilities for security and biology research. Anthropic says the two are the same underlying model; Fable 5.1 ships with safeguards that fall back for cyber and bio capability queries, while Mythos 5.1 keeps those capabilities intact with access limited to approved users.

The stated advance is not primarily about benchmark scores. Anthropic says the more significant improvement is sustained performance in long-horizon workflows: planning multi-step tasks, using tools across applications, adapting when conditions change, and carrying work through to completion. The company also claims better judgment on ambiguous tasks and fewer confident wrong answers, framing this as a behavioral change — the model is said to recognize when it is blocked, report incomplete work, and avoid shortcuts such as resolving a failing test by disabling it.

These are vendor claims, not independent findings, but they target a known failure mode. The harder question for prospective users is whether that behavior holds in production, where the cost of a quiet false success is higher than in a demo.

For teams already on Claude Fable 5, Anthropic says existing prompts should generally carry forward to 5.1, but suggests revisiting prompts built to compensate for earlier model behavior. Repeated warnings not to guess, multiple layers of self-verification, and highly prescriptive completion criteria may no longer be necessary; Anthropic notes that excessive scaffolding can limit a more capable model. It recommends evaluating the new version against representative workloads and simplifying prompts where evidence supports it.

Anthropic also announced Enterprise Frontier Safeguards, positioned as a way for customers to deploy its most capable models while keeping data in cloud infrastructure they control. Enforcement details were not disclosed.

The naming structure matters: Fable and Mythos are the same weights with different safeguard configurations and access controls. For buyers, that means the public-preview name and the restricted research name describe the same model, and the differentiator is governance, not intelligence. For researchers in cyber and bio, access remains gated; for everyone else, the preview version is the relevant one.

Anthropic’s advice is a sensible test: go beyond single-prompt comparisons and try workloads that demand sustained reasoning — a feature spanning a codebase, an agent operating across tools, a multi-step analysis. Whether the model actually says so when it is stuck, rather than reporting success, remains the claim worth checking.