Building the Claude API for the Frontier: effort controls, prompt caching, and Mythos vs Fable safeguards (18:29)

Messages API opinionated defaults and the effort parameter: the same model behaves differently at high and low effort, and the parameter changes output more directly than external prompt engineering.

Feature releases are driven both ways by research and customer requests; changing the system prompt and effort mid-conversation no longer breaks prompt caching prefixes.

Model identifiers serve as versioning boundaries; older capabilities such as fixed thinking budgets are retired through transitional beta headers in favor of adaptive thinking.

Cost management: prompt caching hit rates target up to 90%, and cost is judged by intelligence per token rather than unit price.

Mythos and Fable share the same model weights; Mythos is gated behind a Cyber Verification Program for defensive cybersecurity research, while Fable serves public use with runtime safety classifiers.

Next steps: shelved, sub-threshold API prototypes are re-tested against incoming frontier checkpoints, and failed workflows are re-evaluated as models improve.