Claude Code: choosing the right job for Sonnet, Opus and Haiku

Start with a task you can judge

Claude Code can run a Sonnet session with Opus available as an advisor. Haiku can also take a separately configured exploration task. Those are distinct settings: the command that enables an advisor does not, by itself, create a three-model team.

On current Claude Code releases, the sonnet, opus and haiku aliases currently map to the 5.5 models on the Anthropic API. Other providers can map them differently. Record the actual model versions when comparing runs. Model configuration

Then describe the work precisely. Finding every caller of a function and deciding whether its contract should change demand different kinds of judgment. That distinction is a better starting point for assigning models than a permanent ranking of which one is best.

The useful saving is the work you no longer have to do twice.

Editorial highlight · Original line from this article, not a third-party quotation.

What the advisor command really enables

A documented starting point is:

claude --model sonnet --advisor opus

This selects Sonnet for the main session and Opus as advisor. The official CLI reference does not document a –subagents haiku option, and this command does not configure Haiku workers. CLI reference

The experimental advisor runs on Anthropic’s API infrastructure. Claude chooses when to consult it and sends the conversation and tool activity; those consultations consume additional tokens. Advisor documentation

Treat that arrangement as an experiment with a specific purpose. Perhaps your current workflow spends too long pursuing the wrong explanation for a bug. An additional consultation might help change direction earlier. A smaller bill is only one possible result; better decisions or fewer interventions would also be worth recording.

Give Haiku a bounded assignment

Haiku exploration requires separate configuration. Anthropic documents overriding the built-in Explore agent with a custom agent named Explore and a model: haiku field. Subagents have their own context and return results to the main session. Subagent documentation

Compatible subagents also inherit the advisor setting, so include their consultations in the cost. Advisor documentation

For a first trial, ask for something narrow: locate the files involved in an endpoint, list existing tests or identify references to a renamed symbol. Require paths and enough surrounding evidence to verify the answer. Keep the decision about what to change explicit.

The size of the assignment matters. A tidy summary can still omit the exception that determines the fix. If someone must repeat the entire search to trust it, the delegation has not yet earned its place in the workflow.

Count the finished job

Compare representative tasks using the same acceptance criteria. Record completion time, failed attempts, human corrections and final review. For API-billed work, divide spending across all attempts by the number of accepted tasks, and report the completion rate alongside it. Include unsuccessful runs: a configuration that abandons difficult work can look deceptively economical if you count only its successes.

Claude Code’s /usage screen provides session token statistics and cost estimates; subscription users also have plan-usage information. A displayed API cost estimate is not automatically the subscriber’s bill. Cost documentation

Keep the configuration that gets more changes through review with an acceptable cost and fewer corrections. If the main agent spends its time repairing a subagent’s incomplete search, the cheaper model has simply moved the work elsewhere. Repeat the comparison when model versions change.

Sources and references

• Claude Code CLI reference
• Escalate hard decisions with the advisor tool
• Create custom subagents
• Model configuration
• Manage costs effectively