Newshot
Account

Claude API adds adjustable effort: letting AI scale its effort to the task, shifting gears mid-conversation from casual chat to complex analysis

Source: IT之家 Published: AI 中文 AIClaudeAPI
AI summary

Anthropic has added a Beta feature to the Claude API that lets developers adjust effort within messages, so the same conversation no longer has to use the same setting throughout and developers can change the model's "effort level" mid-conversation based on what later tasks require. Effort can be understood as the setting for how much work Claude puts into answering, helping developers trade off between answer thoroughness, response speed and cost; it affects not just how many words are written at the end but also how much resources the model devotes to thinking, calling tools and organizing its answer. Users can insert an empty system message into the message sequence and set the target level via the new output_config.effort setting, with no need to change top-level configuration and prompt caching still hitting; the setting does not modify content already generated and takes effect from the next user message with new input, so developers need to pay attention to where the adjustment message is inserted and when the next user input appears.

Translation(model translation, not the source's own text)

Claude API Adds Effort Adjustment: Letting AI Apply the Right Amount of Effort per Task, With Mid-Conversation "Gear Shifts" from Chit-Chat to Complex Analysis

IT Home, November 12: Anthropic has added a new Beta feature to the Claude API that lets developers adjust effort within messages. This means the same conversation no longer has to use the same setting throughout; developers can change the model's "level of effort" in its answers mid-conversation according to the needs of subsequent tasks.

IT Home note: effort can be understood as the setting for "how much work" Claude puts into answering a question. It helps developers trade off between how thorough an answer is, how fast it responds, and how much it costs. Effort adjusts more than just how many words are written at the end; it also affects the resources the model puts into thinking, calling tools, and organizing its answer.

In practice, users can insert an empty system message into the message sequence and set the target level via the new output_config.effort. The whole process requires no changes to the top-level configuration, and prompt caching can still hit.

As for when it takes effect, the setting does not modify content that has already been generated; it applies to subsequent answers starting from the next user message with new input. Developers therefore need to pay attention to where the adjustment message is inserted and when the next user input occurs.

In real-world scenarios, a user can set the first half of a conversation for simple Q&A and switch the second half to complex analysis, setting a new effort level at the point where the task changes without having to modify the top-level configuration.

Related Articles

  • Anthropic upgrades its AI office suite: from brainstorming to data dashboards, Claude wants to take over the entire team collaboration workflow
  • "Programming interviews are dead": Anthropic says coding questions can no longer screen programmers in the AI era
  • Anthropic unveils an unreleased AI model: more capable than Mythos 5.1, and more "well-behaved"
  • Anthropic report: AI work hours are already 6 times human hours, junior programmer positions plummet, and revenue per employee is 2 to 4 times that of Meta / Google at the same stage
  • Anthropic test incident: Claude model submitted false murder leads to police and also filed 20 visa applications
  • Anthropic's fastest, cheapest model: Claude Haiku 5.5 arrives, with running costs about 75% lower than 4.5

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗