Anthropic has added the ability to adjust effort for the Claude API.

Anthropic has added a Beta feature to the Claude API to adjust effort, allowing developers to change the 'level of effort' in the model's responses during a conversation. Developers can set the output_config.effort level by inserting an empty system message, affecting resource allocation for model thinking, tool calls, and other aspects, enabling adjustments to the response's comprehensiveness, response speed, and cost during task switching. Effort can be understood as the 'level of effort' Claude puts into answering questions. The entire process does not require changes to the top-level configuration, and prompt caching remains effective. This setting will not modify already generated content but will take effect from the next user message with new input for subsequent responses.

AI-written from the reports · Updated 3 days ago