Anthropic Adds Mid-Conversation Effort Control to the Claude API, So Agents Can Save the Expensive Tokens for the Hard Parts

Anthropic has quietly added one of the more practical cost-control features that agent developers have been asking for: a beta capability in the Claude API that lets them change how hard the model works partway through a conversation. Instead of locking a request into a single effort setting for its entire life, developers can now insert an empty system message carrying a new output_config.effort value, and the model shifts gears from that point onward.

The mechanics are deliberately minimal. The developer places a system message with empty content into the message sequence and sets the target level through the new output_config.effort field. Nothing in the top-level configuration changes, which means prompt cache hits accumulated earlier in the conversation survive the switch. The adjustment is not retroactive: content already generated stays exactly as it was, and the new effort level takes effect starting with the next user message that carries new input. Placement matters, because the switch lands at the boundary between turns, not inside one.

Effort, in Anthropic's design, is not just about how long the final answer is. It governs how much work the model puts into reasoning, how many tool calls it is willing to make, and how carefully it organizes the response. A conversation can start in a cheap, lightweight posture for lookups and clarifications, then move to a higher setting the moment the task turns into genuine analysis, without the developer rebuilding the request or losing the cached prefix.

The economics are the point. Enterprise AI bills have been climbing even as per-token prices fall; reporting by the Financial Times around Nebuly's funding this week noted that enterprise AI spending has roughly tripled for some adopters. Long agent chains are the worst offenders, because they routinely pay premium-model prices for trivial retrieval steps. Mid-conversation effort switching lets a team run the boring 80 percent of a workflow at minimum effort and reserve the expensive posture for the nodes that actually need it, inside one continuous thread.

It also fills an awkward gap in the current toolkit. Switching to a smaller or cheaper model mid-thread usually means handing context across models or losing it, and re-issuing the conversation under new top-level settings risks invalidating the cache. A single model that can be nudged up or down in place keeps the context, the tools and the cached state intact. Alibaba's Qwen4-Max preview documentation exposes a similar reasoning_effort dial with low, medium and xhigh settings, which suggests effort-as-a-runtime-knob is becoming a shared convention rather than an Anthropic quirk.

There are honest caveats. The feature is in beta, and Anthropic has not published benchmark data showing how much money mid-conversation switching actually saves in production workloads. Developers also need discipline about where the switching messages go, since a badly placed adjustment silently applies one turn later than intended, and effort does not substitute for choosing the right model in the first place.

The direction is clear nonetheless. As agents move from demos to payroll-adjacent workloads, the interesting engineering is shifting from the model to the orchestration layer: who decides, per turn, how much thinking the task deserves. Effort controls like this one hand that decision to the application instead of leaving it baked into the API call, and the vendors that make the dial cheap and predictable are the ones enterprises will keep on the bill.