Models, cost and fast mode
The model is the engine a Claude Code session runs on, and choosing it is the first of three dials that decide what a session costs.
Last updated: 28 Sep, 2026 · Claude Code
Three dials change what a session costs: which model, how much it reads, and how much of the conversation it carries. You have already met the second and third.
Choosing a model
/model switches mid-session and --model starts one. Sonnet handles most coding work. Opus is for hard reasoning, such as architecture or a bug no one has explained. Haiku is for volume: the subagent in the subagents lesson searching a repository does not need your best model.
The useful habit is not picking the cheapest model. It is noticing that a lot of the work a session does is not the hard part, and routing that work away from the expensive one, which subagents make easy.
Seeing what it costs
/usageIt prints what has been used against your plan's limits.
The statusline from the statusline and output styles lesson shows the same thing without asking, which is why that lesson exists this early. A number you see all day changes behaviour; a number you have to ask for does not.
For reference, the documentation puts the average across enterprise deployments at around thirteen dollars per developer per active day, with ninety per cent of users under thirty. If you are far above that, the cause is nearly always context rather than the model.
What reduces cost
- Say where to look. A question that names the file skips the searching, from the codebase-search lesson.
- Deny reads of generated code. The last lesson.
- Delegate noisy work to a subagent on a cheap model. The subagents lesson.
- Compact with a focus rather than letting it happen. The context window and compaction lesson.
- Start a new session for a new task. Carrying an unrelated conversation costs on every turn.
Fast mode
/fast is a different trade from all of the above: the same Opus model, configured to answer up to two and a half times faster, at a higher cost per token. It is not a smaller model and the quality is the same. It is in research preview.
Related
- Previous: Large codebases
- Next: Project setup
- Reference: Claude Code docs
- Run
/usageafter a normal working day. - Do the same task twice, once with a question that names the files and once without, and compare the cost in your statusline.
Every expert started right here.