Claude Code effort turns verification depth into a workflow control

Thariq Shihipar reports that higher Claude Code effort levels spend more time and tokens on independent judgment, adversarial checks, and edge-case testing. Across build experiments and Anthropic’s Terminal-Bench 3.0 runs, the gains were strongest when success depended on verification; extra effort was less reliable when the model chose the wrong interpretation or approach. Why it matters: Max effort is not a universal quality switch. It can improve difficult verification while slowing feedback and making more choices on the user’s behalf, so teams should route effort by phase and risk rather than use one setting for an entire workflow.

Try this: Run one feature through a staged effort loop: save the specification and acceptance tests, implement on low or medium, review the diff, then verify on high and compare elapsed time, failures found, and rework.