Sign in Subscribe
news

2026-10-04 Claude effort levels - Draft on low, verify on high

Sebastien Dubois

Sebastien Dubois

— 4 min read

Canonical version: 2026-10-04 Claude effort levels - Draft on low, verify on high.

It's tempting to treat the effort setting in Claude Code as a smartness knob: the task feels hard, so you turn it up. After reading Thariq Shihipar's guide and the community numbers, I think that's the wrong mental model, and an expensive one.

In this piece, I go through what effort buys, where it stops helping, and the settings I'd use for Opus 5.5 and Claude Sonnet 5.5.

What effort buys you

Thariq (Anthropic) published "Spending your effort" on 25 September. He defines effort as roughly how much compute you want Claude to spend, and compares it to giving someone 12 hours for a task instead of 1. In practice, it changes how much verification, edge-case testing and own judgment Claude puts into the task.

His Terminal-Bench 3.0 results (5 attempts per task) show where the extra time goes:

  • Claude Opus 5.5 on mvcc-lsm-compaction: 0/5 at low, 4/5 at xhigh. About 1 minute versus about 11
  • Claude Fable 5.1 on html-js-filter: 1/5 at low, 5/5 at xhigh. On high, it ran about 33 minutes instead of 2. It reviewed its own work adversarially, read the parser's source, ran an XSS test suite and wrote a fuzzer

Everything on that second line is checking: Fable spent the extra 31 minutes testing its own solution.

Where it stops helping

Thariq is explicit about the limit: higher effort reduces edge-case failures, but it does NOT fix failures where the model picked the wrong approach.

Vague prompts also leave more decisions to Claude, and higher effort makes bigger ones. His example is "a personal fitness and workout tracker app": on low, you get a log and a simple graph; on max, something much bigger, with a heat chart. A /config menu redesign took 1 minute on low (an interactive sketch) and 28 minutes on max (a polished mockup with flow walkthroughs). Thariq prefers low first, to see what Claude has in mind before it commits. With a detailed spec, models and effort levels converge.

max can also return nothing at all. Simon Willison asked Opus 5.5 for his pelican-on-a-bicycle SVG at max, and twice it spent the entire 128,000-token budget thinking without producing any output. He called it "virtually useless", and hit the same bug on Sonnet 5.5.

And max isn't always worth the money. Artificial Analysis measured Sonnet 5.5 at $1.08 per task on high and $7.60 on max, which is more than Opus 5.5 at max ($5.98), for a lower score.

In the Hacker News thread on an effort-level analysis of Opus 5.5, one commenter writes that "Medium thinking effort is ideal for most tasks" and finds max prone to overthinking. Another stays on low most of the time and switches models when a task needs more.

Thariq's rule of thumb

  • Low: quick in-the-loop work (brainstorming, sketching, easy changes)
  • Medium: most regular software engineering (feature implementation)
  • High: when verification matters and edge cases lurk (a bug in a brownfield codebase)
  • Max: fully autonomous hard problems (build and verify end to end, security vulnerabilities in critical software)

Effort helps most on hardware, code review and security tasks. The levels are low, medium, high, xhigh and max.

His loop

  1. Write a spec
  2. Ask Claude to interview you about the missing details
  3. Implement on low
  4. Review and iterate on low
  5. Verify and test on high

You only pay for high on the last step, when there's something concrete to check.

Switching mid-session doesn't cost you the cache either. Lydia Hallie (Anthropic): "Switching effort mid-session on Opus 5.5 doesn't break your prompt cache btw! Just make sure you're on Claude Code v2.1.280+". So you can run /effort for the verification pass and drop back down afterward, and your cached context is still there.

Starting points for Sonnet 5.5

The effort levels were recalibrated for Sonnet 5.5, so your old settings don't transfer one to one. Addy Osmani (Anthropic) suggests re-running your effort sweep and starting from:

  • medium for agentic coding (that's also the default in Claude Code)
  • low for chat and extraction

Artificial Analysis calls high its most competitive setting. I wouldn't use max on Sonnet 5.5 at all: it costs more per task than Opus 5.5 at max. If a task needs that much thinking, give it to Opus 5.5.

What I'd do

My defaults:

  • Opus 5.5 on medium, the Claude Code default
  • low to explore: brainstorming, sketches, quick edits, or seeing Claude's take before you commit to a direction
  • high for the verification pass (tests, reviews, edge cases in old code), then back down
  • xhigh and max only for long autonomous runs, with a clear finish line and a token budget you've accepted to burn
  • On Sonnet 5.5, medium by default and high for harder work; switch to Opus instead of reaching for max

When the result is wrong, I'd fix the spec or the approach before touching effort. Raising effort is the last lever I'd pull.

Thariq's full guidance (with the token numbers per level) is in my Claude Code effort levels note. The community analysis is in Claude Opus 5.5, and the Sonnet cost table is in Claude Sonnet 5.5.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe