Choosing Claude models for coding: Anthropic's guide and Claude Code tips

Anthropic published a full guide for choosing between Fable 5.2, Opus 5.5, Sonnet and Haiku. You can ask Fable 5.2 to check your Claude Code configuration against the recommendations from this guide.

Previously, the built-in Explore subagent in Claude Code always ran on Haiku. Now it uses your main model. If that is Opus 5.5 or you have just switched to Fable 5.2, Explore wastes the quota of a more powerful model every time it reads files or searches code.

Therefore, it is worth creating your own subagent named Explore and setting model: haiku.

Prompt to check the configuration

Read this guide: https://claude.dev/blog/which-claude-model-should-you-build-with/

Then check my Claude Code configuration against the recommendations: user and project settings, subagents in ~/.claude/agents and .claude/agents, and the CLAUDE.md file.

Find three categories of problems.

1. Subagents use unsuitable models

Find subagents that only search files, read logs or run tests but do not have an explicit model, so they use the main model.

If there is no custom Explore subagent with model: haiku, treat that as a problem too.

Check the reverse situation as well: subagents that edit code or make decisions should stay on the main model. Make sure the CLAUDE_CODE_SUBAGENT_MODEL variable does not force them to use a weaker model. Check only this variable: in the env block of your settings and in the shell environment.

2. Model or effort level does not match the task

Use the table from the guide:

• Opus 5.5 with medium effort for everyday development and debugging.
• Sonnet 5.5 for clearly defined tasks, such as fixing a bug or implementing a feature from a specification.
• Fable 5.2 for large-scale refactoring, migrations and coordination of many subagents.

Flag effort above high if it is set by default, and suggest trying a more powerful model with a lower effort level first.

Also flag opusplan if it is selected by default, because in this mode code editing runs on Sonnet.

3. There is no way to verify results in the project

Check whether CLAUDE.md says how to run tests and builds, and whether any check can actually be run.

The guide recommends this order: first make sure the result can be checked, then raise the effort level by one step, and only after that move to a more powerful model.

For each issue found

• Quote the relevant setting and the corresponding line from the guide.
• Explain what costs it causes and in which situations it will create problems.
• Propose a minimal fix as a ready configuration that can be copied.

Then study the project code and recent commits. Create a table showing which kinds of tasks in this project are best done on which models and with which effort level.

Finally, explain how to use /model and /usage to run one task from my list on two different models and compare the costs.

Show the recommendations first. Do not change anything yet.