ultrathink is deprecated as of early . /effort replaces it with explicit levels.
low is fastest and costs least. Use it for syntax fixes or simple lookups. medium balances speed and reasoning. high is the default for Sonnet on complex tasks. max applies maximum reasoning depth on your current model.
At /effort max, thinking tokens count as output tokens. You dial it up for hard architectural decisions and dial it down for routine tasks. On a long session with multiple hard problems, running everything at max adds real cost.
I recommend max for decisions you'll live with for months and low for anything you'd fix in under a minute.