How to run Claude Code on Sonnet with Opus as the advisor

Claude Code's advisor lets Claude show the whole conversation to a stronger model at key moments and get guidance back. Sonnet handles the routine turns, and Opus steps in when an approach needs choosing or the same error keeps coming back, and sometimes for a check before Claude declares the task done. It usually costs less than running Opus throughout.

1. Switch the main model to Sonnet and set Opus as the advisor. The choice is saved to advisorModel in your user settings and carries over to later sessions.

/model sonnet
/advisor opus

1. To use it for one session only, pass flags at launch. Your saved setting doesn't change.

claude --model sonnet --advisor opus

1. You can also set it as a default in your settings file.

{
  "advisorModel": "opus"
}

Claude decides when to call the advisor. To make sure it asks, write something like "consult the advisor before you continue" in your prompt. During a call the transcript shows an Advising line, and you can expand the guidance with Ctrl+O.

The advisor must rank at or above the main model: a Sonnet 5.5 main model accepts Opus 5 or later, Fable, or Sonnet 5.5. It is experimental and runs only on the Anthropic API (not on Amazon Bedrock, Google Cloud, Microsoft Foundry or Claude Platform on AWS). It also stays off in sessions where a variable that turns off feature-flag fetching, such as DISABLE_TELEMETRY, is set. Each call has Opus read the whole conversation again, so it adds tokens, and on subscription plans it counts toward your plan limits. Subagents inherit the same advisor, so pairing it with a Haiku subagent for codebase exploration gives you the "Opus advises, Sonnet builds, Haiku investigates" setup going around on X.