Low effort in Claude Code keeps you in the loop

The author of a September 26 guide to effort on Anthropic’s claude.dev developer blog now uses low effort far more often, a level at which Claude asks you for more context where it would otherwise spend tokens working things out alone. If you want to stay in the loop on a task, that is the level the author picks, and max gets little more than two jobs: runs that need zero input and hunts for security vulnerabilities.

Effort, in the post’s terms, gives the model an approximation of how much compute you want it to spend, and it controls how many tokens Claude spends on a response. Summing up the guide on Twitter, the author said they had dug into evals and run their own tests, and the results quite surprised them. Extra effort gave better results in areas where verification and edge-case testing were more useful, such as hardware, code review and security. The benchmarks and demos the author built sit on the claude.dev page as interactive explainers.

The post names responding to effort “without breaking the prompt cache in Claude Code” as one of the best parts of the newest Claude models, and on Claude Fable 5.1, Opus 5.5 and Opus 5 the level can change from one message to the next. A session can stay on low while you steer it and step up to max for a security pass with its cache intact.

In Claude Code, the level is set with /effort inside a session or claude --effort at launch, and the ultracode entry in the /effort menu runs at xhigh, one level below max.