Anthropic Just Revealed Which Effort Level to Use

Summary

The video explains how to choose Claude’s effort level—low, medium, high, ultra-high, or maximum—based on Anthropic guidance.

Effort is not the same as model intelligence. The model choice—such as Opus, Sonnet, or Haiku—has more to do with capability, while effort controls how much work Claude puts into a task. Higher effort generally means more reading, testing, checking, and independent decision-making. Lower effort tends to produce faster results and keeps the user more involved.

A poorly specified prompt cannot be rescued simply by raising effort. Clear instructions and context remain essential. Higher effort can also cost more time and usage, and may lead Claude to make more decisions without asking.

When to Use Each Level

• Low: Use for simple tasks, fast first drafts, brainstorming, labeling, short summaries, rewrites, or work you expect to iterate on. It’s also useful for polishing a result created at a higher level. Because low effort follows instructions closely, give Claude a checklist for tasks with multiple parts. The presenter cites an Anthropic test in which a simple redesign took about one minute at low effort versus 28 minutes at high effort, with a similar result.
• Medium: The recommended default for most everyday tasks where the desired outcome is clear and you’ll review the result once. Examples include routine emails, repurposing content, meeting summaries, recurring reports, and quick research. If a medium-effort answer falls short, first improve the context or instructions before increasing effort.
• High: Use when verification matters, errors could be costly, or the task has many edge cases. Examples include checking financial figures, investigating an unclear trend, reviewing contract risks, fact-checking, and examining long documents. It can also be used as a final inspection after building a draft at lower effort. In one test described in the video, low effort tried one approach to a data-analysis task and got it wrong, while high effort compared approaches and got the correct result four times out of five.
• Ultra-high or maximum: Reserve these for long, difficult, critical tasks when you plan to leave Claude working without supervision. Examples include extensive research across a year of support data, reviewing a consequential agreement, or producing a high-stakes presentation from several source documents. The presenter says ultra-high is generally the more economical starting point, while maximum is rarely needed. Anthropic documentation is described as warning that maximum can add substantial overhead for relatively small quality gains—and can sometimes lead to overthinking or worse results.

The presenter also recommends a staged workflow: create and refine a draft at low effort, then switch to high effort for thorough testing or verification. He says newer models can switch effort levels mid-conversation without the earlier full-cost rereading penalty.

Quick Selection Guide

Ask these questions in order:

1. Is the task simple, or will you actively iterate with Claude? Use low.
2. Could a missed error cost money, a customer, or lead to a bad decision? Use high.
3. Is the task long, difficult, and critical—and will you leave Claude to work independently? Use ultra-high or maximum.
4. Otherwise, use medium. If uncertain, medium is the suggested default.

Guide and Tool Featured

The video provides a practical guide to selecting effort levels, with examples for each. It also promotes a free Claude Code mod that analyzes prompts and suggests a different effort level when it detects a likely mismatch. The presenter says it can offer a one-click change or apply the suggestion automatically.

Main Speakers and Sources

• Ben, the video’s presenter.
• Thariq Shihipar and Lydia Hallie, Anthropic team members whose guidance is discussed.
• Anthropic research and documentation, cited for effort-level recommendations and test results.