Model and Effort in Claude Code: Anthropic Explains Two Often-Confused Settings

Anthropic clarifies Claude Code by distinguishing its model setting, which sets capacity, from its effort level, which governs the steps taken on a query.

Two parameters in Claude Code seem to play the same role in improving the response: model choice and effort level. Anthropic published a post to clear up the confusion. The first setting fixes the model's weights, thus its capacity and knowledge base, which are set during training and are read-only at inference time: the provided context guides the prediction without ever modifying these weights. The second, effort, is not limited to thinking time: it governs the amount of work engaged on a query, meaning the number of files read, checks performed, and steps taken before the model returns to the user.

At high effort, the agent reads more, tests, cross-references, and pushes a multi-step task further before handing it over; at low effort, it prefers to ask for more context rather than spending tokens to decide on its own. Anthropic specifies that effort is not capped by this slider, only maxtokens truncates an ongoing response, and that its teams monitor overthinking during training, an excess of zeal that degrades efficiency on simple tasks.

The recommendation boils down to a heuristic. Stick to each model's default effort, adjusted as a general preference based on the type of work rather than on a case-by-case basis. When encountering an error, the author suggests first checking the provided context, then asking a question: did the model not know enough, or did it not try hard enough? If the context was complete and the model clearly tried but failed, increase the model's capacity; if it rushed, skipping a file or abandoning a refactor halfway through, increase the effort level.