Which Claude model should you use?
Claude comes in tiers, and the best choice depends on the task. A fast model is cheaper and quicker. A larger model solves harder problems but costs more and takes longer.
The short answer#
- Start with Sonnet. It is a strong everyday default for coding.
- Switch to Opus when a task is hard: a tricky bug, a large refactor or a design decision.
- Use Haiku for quick, simple or repetitive work where speed and cost matter most.
The three tiers#
Haiku: fast and cheap
Best for renaming, formatting, simple edits, classification and quick questions. It is also a good fit for helper agents that run many times.
Sonnet: balanced
Best for most feature work, bug fixes, tests and code review. It gives a good mix of quality, speed and cost.
Opus: most capable
Best for hard reasoning, planning large changes and problems where earlier attempts failed. Expect slower answers and higher cost.
The current Claude 5 family#
These are the model IDs for the current family. Use them when you call the API or set a default model:
claude-haiku-5-5
claude-sonnet-5-5
claude-opus-5-5
claude-fable-5-2Capabilities, limits and prices change with each release, so this guide does not list benchmark scores that would soon be out of date. Check Anthropic's model documentation for the latest details, including what Fable 5.2 is best at.
Switch models in Claude Code#
# inside a session
/model
# or choose when you start
claude --model opusYou can also set a default in .claude/settings.json with the model setting, or pin a model on a single subagent in its frontmatter.
A fair way to decide#
- Pick a real task from your project.
- Run it on two tiers, for example Sonnet and Opus.
- Compare the quality of the result, the time it took and the cost.
- Use the cheaper tier whenever the quality is good enough.
Frequently asked questions
Is a bigger model always better?
No. For simple tasks a smaller model is often just as good, and faster and cheaper.
Can I use different models for different jobs?
Yes. You can switch with /model, and subagents can each use their own model.