Use medium, Claude Opus 5.5's default since its September 22, 2026 release, for most founder thinking, and step up to high or xhigh only for decisions that need Kahneman's outside view or a premortem. Artificial Analysis data shows max costs $5.98 per task against $1.34 at medium for 6.4 more index points, and Thinking, Fast and Slow explains why: longer deliberation makes the story more coherent, not more correct.
Anthropic released Claude Opus 5.5 on September 22, 2026, and the change that matters most for anyone using it to think, rather than to code, is the one buried in the docs: it defaults to medium effort, and effort is now the main dial for how much the model reasons and what a request costs. I spent three days running my own decision questions through all five levels, and the result is less about the model than about a bias Daniel Kahneman described fifteen years ago.
The specifics, because they change the calculation. Opus 5.5 is priced at $4 per million input tokens and $20 per million output, with a 1M-token context window and 128K maximum output; Anthropic says it performs at the level of Claude Fable 5.1, which costs $10 and $50, on most work, and costs 40% less than Opus 5 at default settings. Adaptive thinking is always on and cannot be disabled, so the five effort levels, low, medium, high, xhigh and max, are the only control, and because Opus 5 defaulted to high, a request that omits the setting now runs one level lower than it did last week. What each level costs in practice comes from Artificial Analysis measurements on its Intelligence Index, compiled by FinOps LLM on September 22: low scores 42.3 for $0.55 per task on about 10,200 output tokens; medium 51.2 for $1.34 on 25,700; high 53.6 for $1.82 on 35,600; xhigh 56.0 for $3.46 on 65,700; and max 57.6 for $5.98 on roughly 119,200 output tokens. Going from xhigh to max buys 1.6 points for 73% more money. Anthropic's own headline comparison against GPT-6 Astra is made at medium. And on Hacker News this week one user reported that at max the model twice ran out of its 128,000-token budget while still reasoning about a simple SVG drawing.
The lens I used is Thinking, Fast and Slow, because effort is literally a System 2 budget. Kahneman's System 2 is slow, effortful and lazy; it engages when cognitive strain demands it and otherwise endorses whatever coherent story System 1 has already built. The trap is that more deliberation does not reliably mean more accuracy. "Subjective confidence in a judgment is not a reasoned evaluation of the probability that this judgment is correct," he writes; confidence tracks the coherence of the story, what he calls WYSIATI, and a long reasoning trace is a very coherent story. Where deliberation does pay, in Kahneman's account, is in specific moves: taking the outside view by asking for base rates, running a premortem that assumes the decision failed, and replacing judgment with a simple formula where one exists. So the question for each effort level is not "how smart is it" but "which of those moves does the extra spend actually buy."
| Effort level | AA index / cost per task | Kahneman test it passed in my runs | Founder job I assign it |
|---|---|---|---|
| low | 42.3 / $0.55 | Cognitive ease only: fluent, no checking | Summaries, drafts, reformatting |
| medium (default) | 51.2 / $1.34 | Coherent story plus one caveat when asked | Weekly metrics read-through, first-pass memos |
| high | 53.6 / $1.82 | Outside view when I ask for base rates | Hiring and vendor decisions |
| xhigh | 56.0 / $3.46 | Unprompted premortem with failure modes I missed | Pricing changes, fundraising plans, anything irreversible |
| max | 57.6 / $5.98 | Re-argued both sides, ended less decisive | Nothing yet |
What that table looked like from the inside. My test question was whether to drop the annual plan for a product I advise. At medium, Opus 5.5 gave me a one-screen recommendation with a clean argument, exactly the coherent story Kahneman warns will feel true. At high, once I asked for base rates on annual-to-monthly migrations, it pulled the discussion toward churn timing instead of price, which was right. At xhigh, without being asked, it wrote a premortem that named two failure modes I had not considered, the sales team's incentive to discount and the cash-flow gap in month two, and that is the moment the extra $1.64 earned itself. At max it produced roughly 4,000 words that argued both sides twice and ended less decisive than xhigh, which is Kahneman's illusion of validity in a new costume: the story got more elaborate, not more right. The most useful finding was cheaper than any of this. Asking explicitly for the outside view and a premortem at high got me most of xhigh's value for about half the cost. The setting to change first is the question.
Limits. These are a few dozen prompts on one founder's questions, not a benchmark, and the Artificial Analysis index measures a mixed suite, not decision quality. Anthropic's docs say effort is "a behavioral signal, not a strict token budget," so the same level thinks less on an easy question and more on a hard one, and the model has already changed its style between versions, so rerun your own sweep rather than trusting mine. The costs are API prices; in the consumer app the dial is exposed differently and a subscription user mostly gets the default. And the deeper failure mode is not the effort level at all. It is treating a longer answer as a more considered one, which is the thing System 2 was supposed to protect us from.
Sources: Anthropic, "Introducing Claude Opus 5.5," September 22, 2026 (anthropic.com/claude-opus-5-5); Claude Platform Docs, "Claude Opus 5.5" model overview with pricing, context and default effort (platform.claude.com/docs/en/models/opus-5-5/overview); Claude Platform Docs, "Effort," effort-level table and Opus 5.5 recommendations (platform.claude.com/docs/en/build-with-claude/effort); FinOps LLM, "Opus 5.5 effort levels are the real price," September 22, 2026, compiling Artificial Analysis Intelligence Index v4.3.2 cost-per-task data (finopsllm.com/research/opus-5-5-effort-level-cost); Hacker News thread, "Claude Opus 5.5 Intelligence, Performance and Price Analysis (Max)," September 2026 (news.ycombinator.com/item?id=49804316); Daniel Kahneman, Thinking, Fast and Slow (Farrar, Straus and Giroux, 2011).
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · Why Exploration Is Important for Success