Anthropic launches Claude Fable 5 and restricted Mythos 5

Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9. The release post says the products use the same underlying model but apply different safeguard and access policies. Fable 5 was generally available through Anthropic’s API and products, while Mythos 5 initially served a restricted group of Project Glasswing cyber defenders and infrastructure partners.

Safeguards became part of the model contract

Anthropic says Fable 5 routes some cyber and biology requests to Claude Opus 4.8 when classifiers intervene. It acknowledged that conservative tuning would catch some harmless requests. Mythos 5 removed selected safeguards for trusted-access work, making organization eligibility and use-case controls part of deployment rather than a simple feature flag.

Applications therefore need to record more than the family name. Capture requested model, effective fallback model, safeguard event, access program, and policy version. If an output must be reproducible, a later run under a different classifier can be materially different even when the underlying model identifier appears unchanged.

Evaluate useful completion and intervention together

The launch includes extensive Anthropic benchmarks and early-customer examples across coding, vision, cyber, and life sciences. Those results are vendor-reported. A production evaluation should measure accepted task completion alongside refusal, fallback, latency, token consumption, and reviewer correction. Cyber and scientific suites should include both clearly benign work and boundary cases permitted by the organization’s policy.

Pin the explicit API model during evaluation and create a user-visible path for fallback. A rerouted answer should not masquerade as the originally requested model. Before rollout, verify what happens to streamed responses, tool calls, cached prompts, and cost accounting when a safeguard changes the effective model mid-workflow.