Claude API adds mid-conversation effort switching in beta

According to an IT Home report on November 12, 2026, Anthropic has added a Beta feature to the Claude API that lets developers adjust the effort level in the middle of a conversation instead of sticking to a single top-level setting. In the report, effort is described as how much work Claude puts into an answer, giving developers a way to trade off thoroughness against response speed and cost. The setting does not just control how many words the final reply contains; it also affects the resources the model spends on reasoning, tool calls, and organizing its response. Operationally, a developer inserts an empty system message into the message sequence and sets the target level through a new output_config.effort field. No change to top-level configuration is required, and prompt caching still applies. As for timing, the new setting does not alter content that has already been generated. It takes effect on subsequent answers starting from the next user message that carries new input, so the placement of the inserted message and the moment of the next user turn matter. The report gives a practical example: the first half of a conversation can handle simple questions, and when the later part shifts to complex analysis, a new effort level can be set at the task boundary rather than reworking the top-level configuration. The feature remains in Beta in this report, and no later rollout or version milestones are mentioned.