Within the same conversation, it can start loose and then tighten up—Anthropic is finally letting developers give the model a mid-stream "gas pedal." On November 12, the company behind Claude added a Beta capability to its API: developers can adjust effort at any point in the message stream—that is, the setting for how much "effort" the model is willing to spend answering—helping them easily balance answer completeness, response speed, and usage cost.

Shift gears without restarting, and the cache stays online

Previously, a conversation could only use one fixed set of parameters; now that is no longer necessary. The method is to insert a blank system message into the message sequence, then use the new output_config.effort to specify the target level, while the top-level configuration and any prompt cache hits already made are preserved as-is. Keep in mind that this adjustment is not retroactive—it only takes effect starting with the next user message containing new input. Content already generated remains completely untouched, so where to insert it and when the next question will arrive both need to be calculated precisely.

From chit-chat to hardcore, one dividing line

effort controls far more than the final word count; depth of thinking, the number of tool calls, and the effort put into organizing the answer are all affected. A typical use is: do lightweight Q&A in the first half of the conversation, then turn the level up at the moment you switch to complex analysis, and the model automatically applies more effort. It is especially cost-effective for long-chain agents—use a low level for early retrieval to save tokens, then shift up at nodes that truly require reasoning, without having to pay upfront for the entire conversation. When AI learns to "cut its coat according to its cloth," compute power and the bill are truly in developers' hands.