The advisor tool: a cheap model does the work, a strong model advises at key moments

Running Opus on every turn is expensive; running Sonnet or Haiku is cheap but occasionally picks the wrong direction at key decisions, and the rest of the task pays for it. The advisor tool offers a combination: the main model (cheap, fast) does most of the work, and at important moments Claude consults a stronger model. The advisor reads the whole conversation, including every tool call and result, and returns guidance for Claude to apply. The feature is experimental.

When does Claude consult the advisor?

Claude decides the timing, usually:

• Before committing to an approach.
• When an error keeps recurring.
• Before declaring a task done.

There's no setting to force the number of calls; if you want Claude to consult more or less, say so in the prompt, e.g. "consult the advisor before continuing".

Enabling the advisor: three ways

/advisor opus             # in a session, saved as the default
claude --advisor opus     # for one session only

// settings.json
{ "advisorModel": "opus" }

Turn it off with /advisor off. During a consultation the transcript shows an Advising line; press Ctrl+O to read the advisor's guidance.

Suggested model pairings

The rule: the advisor must rank at or above the main model. If you choose a weaker advisor, Claude Code doesn't attach it to the main model's requests.

Cost

• Each call, the advisor reads the whole conversation at the advisor model's rates, on top of the main model's usage.
• Subscriptions: counts toward plan limits. API: billed at the advisor model's input/output rates.
• The advisor's read of the conversation is not cached; each call processes the full transcript again. So the longer the conversation, the more each consultation costs.
• Toggling the advisor mid-session doesn't invalidate the main model's prompt cache.

Because Claude only consults at decision points, pairing a fast main model with a strong advisor is usually cheaper than running the strong model all session.

Compared with other ways to combine models

Limits

• Anthropic API only; not on Bedrock, Vertex (Google Cloud) or Foundry. Through a gateway, it depends on whether the gateway forwards the request intact.
• Needs feature flags; setting variables such as DISABLE_TELEMETRY turns the advisor off.
• Claude usually follows the advisor, but if real evidence contradicts it (a suggested step fails), it surfaces the conflict instead of following blindly.

How to apply it

• Dev: Try Sonnet + Opus advisor for large refactors and debugging sessions with recurring errors; compare /usage with running Opus all session.
• Tech Lead: Set a default advisorModel for teams using a cheaper main model, to keep planning quality without inflating cost.
• PM: The advisor is a way to raise quality with controlled cost; measure it by the amount of rework, not just the bill.