Anthropic has launched a beta feature for the Claude API: developers can adjust the effort tier at any point in the message stream to balance speed, thoroughness, and cost, ideal for agent scenarios.Author and source: AIBase
In the same conversation, shifting from relaxed to focused, Anthropic has finally enabled developers to "step on the gas" mid-stream. On November 12, this Claude developer introduced a beta feature to its API: developers can now adjust the effort level at any point during the message stream, allowing them to fine-tune the model’s willingness to "put in effort," helping users balance answer depth, response speed, and usage cost with ease.
Switch without restarting, cache remains connected
Previously, each conversation could only use a fixed set of parameters, but that’s no longer necessary. The approach is to insert a blank system message into the message sequence and then specify the target level using the new output_config.effort, while keeping the top-level configuration and any previously matched prompt caches unchanged. Note that this adjustment is not retroactive—it only takes effect starting with the next user message containing new input, leaving all previously generated content untouched. Therefore, you must carefully plan where to insert it and when to wait for the next query.
From casual chat to hardcore— a turning point
Effort extends far beyond the final word count—it also involves depth of thought, frequency of tool usage, and investment in organizing responses. A typical approach is: use lightweight Q&A in the first half of the conversation, then elevate the level precisely when complex analysis is needed, prompting the model to automatically increase its effort. This is especially cost-effective for long-chain agents—use low-tier settings for initial retrieval to save tokens, then upgrade only when genuine reasoning is required, avoiding upfront costs for the entire dialogue. When AI learns to “adjust effort according to need,” developers truly gain control over both compute resources and billing.