Claude Code Advisor: a Second Opinion for Your AI

The Claude Code advisor pairs your main model with a second, typically stronger model that Claude consults at key moments: before committing to an approach, when an error keeps recurring, and before declaring a task done. You enable it with /advisor, pick the advisor model, and Claude decides when to ask for help.

I'm Evgeny, and here's how I explain it to non-programmers: you've hired a capable junior who works fast and cheap, and you've given them a senior architect's phone number. They don't call about every semicolon — they call before big decisions. That's the advisor strategy, and it's one of the smartest cost tricks in Claude Code right now.

What is the advisor strategy, exactly?

When Claude calls the advisor, the second model receives the full conversation — every tool call and result included — and returns guidance that Claude applies before continuing. It runs server-side on Anthropic's infrastructure, works for both subscription and API-billed accounts, and the timing is model-driven: Claude tends to consult before committing to a plan, when stuck on a repeating error, and before claiming completion. It's experimental, requires Claude Code v2.1.98 or later on the Anthropic API, and isn't available on Bedrock, Vertex, or Foundry.

How do I turn the advisor on?

Three ways, depending on how committed you are. The /advisor command sets it mid-session and saves it as your default. The advisorModel key in your settings file makes it permanent without opening a session. The --advisor flag enables it for one session only and wins over the saved setting.

/advisor opus

Which model should be the advisor?

One rule: the advisor must be at least as capable as the main model. Haiku and Sonnet mains can take Fable, Opus, or Sonnet as advisors; an Opus main needs Fable or an Opus at the same version or above; a Fable main accepts only Fable. The pairing I recommend to beginners is the docs' classic: Sonnet as the main model, Opus as the advisor — Sonnet does the routine work, Opus weighs in on plans, weird failures, and completion checks. The thriftiest option is Haiku main + Opus advisor: cheapest day-to-day, with strong planning on call. Note the pairing is enforced by the API, not by the menu — an invalid combo saves fine, then errors on the next request.

How much does the advisor cost?

Each consultation sends the whole conversation to the advisor model, so those tokens are billed at the advisor's rates on top of your main model's usage — and on subscriptions they count toward your plan limits. But Claude calls it at decision points, not every turn, so a fast main model plus a strong advisor typically costs less than running the strong model throughout. Check the damage anytime with /usage. Bonus detail from the docs: toggling /advisor mid-session does not invalidate your prompt cache, so flipping it on costs nothing by itself.

How do I see what the advisor said — and turn it off?

During a consultation the transcript shows an Advising line with the advisor's name; press Ctrl+O to expand it and read the full guidance. Claude generally follows the advice, but if its own evidence contradicts a specific claim — a recommended step fails, a file says otherwise — it surfaces the conflict instead of obeying blindly. To stop using the advisor, run /advisor off or pick No advisor in the picker; that clears your saved advisorModel.

/advisor off

The kicker: next time Claude says a task is finished and you feel that tiny doubt, don't argue with it yourself. Type "consult the advisor: is this really complete?" — and let the senior architect take the call. That's what their phone number is for.