Start at medium, the default on Claude Opus 5.5, and raise it one level only when an answer fails for lack of reasoning. Artificial Analysis data shows max uses about 4.5 times medium's tokens to solve 58 of 100 hard tasks versus 51. Kahneman's Thinking, Fast and Slow explains the trap: importance is not difficulty, and more thinking cannot fix missing facts.
Two of the questions people type about Claude this autumn are "opus 5.5 effort level" and "opus 5.5 thinking level". They are the same question, and the short answer is that there is now only one dial. The longer answer is about how founders choose its setting, because most of us pick it by the wrong criterion.
Claude Opus 5.5 has five effort levels: low, medium, high, xhigh and max. According to Anthropic's effort documentation, medium is the default on Opus 5.5, while Opus 5 and most other Claude models default to high, so a request that sets nothing runs one level lower than it did on the previous Opus. Thinking is no longer a separate switch: Claude's help center says thinking cannot be turned off on Opus 5.5 and is on at every effort level, so to make it think less you lower effort. In the Claude apps the effort selector sits in the model menu next to the send button; in Claude Code you type /effort followed by the level. Higher effort uses more tokens, which means you reach your plan's usage limits sooner.
The cost and benefit numbers come from Artificial Analysis, which runs Opus 5.5 at every level. With medium as the baseline, Pascal Reischl's September 2026 summary of that data shows low using about 0.4 times the tokens per task and solving 42 of 100 Intelligence Index tasks, medium solving 51, high 54 at 1.4 times the usage, xhigh 56 at 2.6 times, and max 58 at 4.5 times. At max, Opus 5.5 writes around 119,000 output tokens per test task, most of it thinking. The curve flattens quickly: going from high to xhigh nearly doubles usage for two extra solved tasks out of a hundred. Anthropic's own documentation says max adds significant cost for relatively small quality gains on most workloads, and can lead to overthinking on structured or less intelligence-sensitive tasks.
Daniel Kahneman's Thinking, Fast and Slow explains why founders still reach for max. We confuse importance with difficulty. A board memo or a pricing decision feels weighty, and the feeling of weight triggers the instinct to throw maximum effort at it. But Kahneman's System 2, the slow and deliberate mode, only helps where a problem actually fails for lack of reasoning. His WYSIATI rule, what you see is all there is, names the bigger trap: a mind builds the most coherent story it can from the information in front of it, without asking what is missing. A model does the same. If your memo leaves out the churn numbers, max effort produces a longer, more confident answer built on the same gap. More thinking cannot supply facts you never gave it.
So the useful question is not "how much does this matter?" but "why would this answer be wrong?" That turns into a short routine.
The difficulty-not-importance check (five steps)
- Start every serious decision at medium. It is the default, and Artificial Analysis places it on the cost-performance frontier along with high, xhigh and max.
- Before touching the dial, read the answer and name the gap. If what is missing is context, such as a number, a constraint or a customer quote, fix the prompt. Raising effort will only polish the gap.
- If the gap is reasoning, raise one level and rerun. A step it skipped, a calculation it fudged, an option it never weighed. Compare the two answers side by side; if you cannot find the difference, go back down.
- Keep xhigh for long multi-step agent runs, which Anthropic describes as tasks over 30 minutes, and use max only after xhigh has failed on a genuinely hard problem.
- Spend the budget you saved on a second pass from the opposite side. For a real decision, a medium-effort premortem ("assume this failed a year from now; what went wrong?") usually does more than one max-effort answer, because it changes what the model looks at rather than how long it looks.
The limits of this advice are real. The Intelligence Index tasks are expert-level problems in science, coding and office work; on everyday founder tasks the levels sit closer together in usage, so treat the ratios as a guide to direction, not a price list. Anthropic says effort is a behavioural signal rather than a strict token budget, and it does not publish how much plan usage each level costs, so these figures are approximations from API-priced benchmarks. There are genuine cases for max: a dense legal or financial document where one missed clause is expensive, or a hard technical diagnosis that has already defeated xhigh. The point is to arrive there because a lower level demonstrably missed something, not because the task felt big. For choosing between models in the first place, see GPT-6.1 Terra vs Claude Sonnet 5.5 vs Opus 5.5 for strategic thinking, and for the cost side of heavy use, what a $200K token budget buys.
Sources: Effort, Claude Platform Docs, Anthropic, 2026; Change the model, effort, and thinking settings, Claude Help Center, November 2026; Claude Opus 5.5 takes the top spot on the Artificial Analysis Intelligence Index, Artificial Analysis, September 2026; Your Claude plan isn't too small. It's set to the wrong level., Pascal Reischl, September 27, 2026.
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours