A small update for Claude developers, but one that people working on agents will probably appreciate.
Anthropic has added a beta capability to the Claude API: in the same conversation, you can adjust the model's effort level at any time, that is, how much effort it is willing to spend when answering questions.
Previously, effort was the top-level parameter of the entire request, and a session used one level from beginning to end. If you want to change it midway, you must either restart the session or change the top-level configuration. The latter will invalidate the hit prompt cache, and the cost will increase immediately.
The new approach is to insert a system message with empty content in the message sequence, and use output_config.effort to specify the new level. The top-level configuration is unchanged, and the cache is still hit.
There is one detail that needs to be figured out: the new gear does not backtrack. It takes effect from the next user message with new input, already generated content is not affected.
Effort affects not only the length of the answer, but also the depth of thinking and the number of calls to tools. Typical usage is: use low setting in the previous stages of retrieving and organizing data to save tokens and run faster; turn it high when it comes to the point where reasoning and decision-making are really needed. In long-link agent tasks, it is much cheaper to adjust gears in sections than to switch to high gear throughout the entire process.
When using it, bring the beta header mid-conversation-output-config-2026-07-01. Supports Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5 and other models, and can also be used on Amazon Bedrock, Google Cloud, and Azure.