Anthropic published the actual diagnostic for model tier versus effort level.

Official guidance separates two knobs people conflate constantly. Effort is not thinking time. It governs how many files Claude reads, how many tools it calls, and how many steps it takes before checking back. The rule: if Claude had all the context, clearly tried, and was still wrong, raise the model tier. If it skipped a file, never ran the tests, or bailed mid-refactor, raise effort. Their stated first instinct for either failure is to fix the context you supplied, not to turn a knob. I've been raising the model tier for what were plainly effort failures for months.