The variable everyone skips
Pick any two Anthropic models and the gap between them is smaller than the gap between one model at minimum effort and the same model at maximum. That single fact reframes the whole question.
On Cursor's own benchmark, Opus 5.5 Low scores 43.7 percent at $1.17 a task. Opus 5.5 Max scores 57.8 percent at $13.43. Same weights, same training, 14.1 points apart, and roughly 11 times the money.
So "which model" is the second question. The first is how hard you are asking it to think.
Where the returns stop
The Opus 5.5 rows are unusually instructive because High and Extra High land on the same score. Cost does not: Extra High costs roughly 1.8 times what High does for a result the benchmark cannot distinguish.
That is what a saturated dial looks like. Past a certain point, more thinking produces longer transcripts rather than better answers, and the extra steps are the model second-guessing itself. High is the sensible ceiling for routine work; Max is for the specific task that has already defeated everything else.
When the heavier model actually earns it
Three situations, and they share a shape: the cost of a wrong answer is high and the problem is genuinely underspecified.
- Ambiguous multi-file refactors. Where the right change depends on reading intent across several files rather than applying a rule. This is precisely what CursorBench measures, which is why the ordering there is worth trusting for this case specifically.
- A bug that survived two attempts. If a lighter model has already failed twice, a third attempt from it is unlikely to differ. The escalation is cheaper than your time.
- Design decisions you will live with. Choosing an abstraction, a schema, a migration strategy. The token cost is trivial against the cost of unwinding it in three months.
Everything else, meaning the large majority of agent turns, is a rename, an import fix, a test stub or a file lookup. Paying frontier rates for those is how an allowance disappears by the middle of the month.
The practical pattern
Leave the picker on Auto, which draws from the cheaper Cursor Models pool entirely. When a task resists, switch to an Anthropic model at a middle effort for that task, then switch back. Escalating per task rather than per month is the difference between a Pro plan that lasts and one that does not.
Model list from the models and pricing table, scores from CursorBench, read on 25 September 2026. Benchmarks move; check both before making a spending decision.