Advisor: Opus only when it counts

What it is and why it matters

You use Opus for everything and by midweek your plan's limit is gone. But almost all of a session is routine: reading files, changing two lines, running the tests. Opus matters in only a few moments: when you pick the approach, when an error keeps coming back, before you say you're done.

That's what the advisor does. You work with a lighter model, for example Sonnet, and pick a stronger model as its consultant. At key moments Claude consults it: the advisor reads the whole conversation, tool calls included, and gives guidance before Claude carries on. Claude decides when to call it.

It's for long sessions: a big refactor (rewriting code without changing what it does), a bug that keeps returning, a job to double-check before you close it. On short tasks it adds little. And it's experimental: behavior, pricing and availability may change.

1. What you need, and where it doesn't work

If one of these three things is missing, the advisor stays off:

• The Anthropic API, directly or through a subscription. It isn't available on Amazon Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform or Microsoft Foundry.
• A recent main model: Fable, Opus, Sonnet or Haiku. And Claude Code kept up to date.
• Feature flags, the switches Claude Code downloads from Anthropic: with DISABLE_TELEMETRY or a similar variable the advisor doesn't turn on.

To remove it completely there's CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1: the /advisor command disappears and advisorModel is ignored.

2. The three ways to turn it on

When it's active, at the start of the session you see the notice Advisor Tool (experimental) is on and may use more tokens.

Way 1: the /advisor command

Type /advisor on its own and a menu opens with the models that can act as consultant and a No advisor option. Or type the model straight away:

/advisor opus

It answers Advisor set to and the model name, and saves your choice to advisorModel in your user settings: it carries over to later sessions. To turn it off:

/advisor off

Way 2: advisorModel in your settings

To have it as your default without opening a session, write it in ~/.claude/settings.json. If the file already has other settings, just add the line, with the comma:

{
  "advisorModel": "opus"
}

Way 3: the --advisor flag, for a single session

To try it without touching your default, launch Claude Code like this. It applies to that session only and takes precedence over advisorModel. If the main model doesn't support the advisor, or the model you ask for can't act as a consultant, it exits straight away with an error.

claude --advisor opus

3. The six model pairings, and when to use them

The rule: the advisor must rank at the same level as the main model or above it. For example, with the latest version of Opus as the main model, only Fable or an Opus of the same generation will do. If you pick a weaker advisor, Claude Code doesn't use it and tells you with a notice.

The six pairings from the official documentation:

1. Sonnet + Opus as advisor: the pairing to start with. Sonnet does the routine work and hands Opus the planning, the unclear failures and the final check.
2. Sonnet + Fable as advisor: Fable's guidance at decision points, without running Fable for the whole session. Requires Fable access.
3. Haiku + Opus as advisor: the cheapest main model with strong planning. It costs more than Haiku alone, but less than switching to Sonnet or Opus.
4. Opus + Opus as advisor: a second Opus reviews the first. For high-stakes work, where an independent check matters more than cost.
5. Fable + Fable as advisor: the most capable pairing. Requires Fable access, and with Fable as the main model Claude Code won't apply an Opus or Sonnet advisor.
6. Sonnet + Sonnet as advisor: a low-cost second opinion for routine oversights.

For Fable, check /model: on the Anthropic API it's there unless your organization excludes it. On some plans Fable bills to usage credits, as an advisor too, and you first accept a one-time consent with /model fable. That also makes Fable your main model, though: go back to yours with /model sonnet, then type /advisor fable.

4. When it calls the advisor and what you see

Claude usually consults it before committing to an approach, when an error recurs and before declaring a task done. There's no setting to cap or force it: if you want it more or less often, say so in the prompt.

During the consultation you see an Advising line with the consultant's name. Then the line tells you how it went:

• Reviewed: the advisor has read the conversation. If there's readable guidance, open it with Ctrl+O.
• Declined: the advisor chose not to give guidance. The reason, if there is one, you read with Ctrl+O.
• Unavailable: the call failed, and the line shows the error code.

Claude follows the guidance, but not blindly: if a recommended step fails or the files say otherwise, it shows you the conflict.

5. What it really costs

Every call makes the advisor read the whole conversation, at the consultant model's rates and on top of the main model. With the API you pay for those tokens; on a subscription they count toward your plan's usage limits, except Fable on plans where it bills to usage credits. It typically costs less than running the strong model for the whole session, but more than the light model alone. You see it in /usage.

Turning it on or off mid-session doesn't throw away the main model's cache, the part of the conversation already stored that makes later requests cheaper. Switching models with /model does.

Anthropic's numbers, measured on the API and not inside Claude Code (The advisor strategy, April 9, 2026, claude.com/blog/the-advisor-strategy):

• Sonnet with an Opus advisor: 2.7 points more on SWE-bench Multilingual, a programming test, and a cost per task 11.9% lower than Sonnet alone.
• Haiku with an Opus advisor on BrowseComp, a web research test: 41.2%, against 19.7% alone.
• Haiku with Opus trails Sonnet alone by 29% in score, but costs 85% less per task.

6. Prompts to copy

In the Claude Code chat, with the advisor on. Replace the parts in [ ] with your own.

Before picking the approach

Before writing any code, consult the advisor on the plan. I want to [add Google sign-in]: suggest two approaches, ask the advisor which one holds up better and tell me what it advised.

When an error keeps coming back

This error has come back three times: [paste the error]. Before another fix, consult the advisor and tell me what it thinks the real cause is.

The session plan for a long refactor

Refactor [src/api] without changing what the code does.
1. Read the code, write the plan in PLAN.md and consult the advisor on the plan.
2. One file at a time, with the tests after each one.
3. If a test fails twice for the same reason, stop and consult the advisor.
4. No advisor for routine edits.
5. Before saying you're done, consult the advisor for the final check and tell me what it found.

Common mistakes

• Turning it on for a ten-minute job: /model is enough.
• Giving consent with /model fable and staying on Fable without meaning to.

Checklist

• At startup I see the Advisor Tool (experimental) is on notice
• My advisor ranks at the same level as my main model or above
• I saw an Advising line and read the guidance with Ctrl+O
• After the session I checked in /usage how much the advisor used

Frequently asked questions