Anthropic Put an Advisor Inside Claude Code. It's Called /advisor.

Anthropic just introduced a new command in Claude Code that gives the assistant a backup brain for the hard moments. You pick a fast, cheap primary model to do the work, and set a second, stronger model that only steps in when things get tight.

How it works

The primary model (the executor) runs the task end to end: reads files, calls tools, iterates. When it hits a decision it can't resolve alone, whether before picking an approach, facing a repeating error, or before marking a task done, it consults the advisor. The advisor receives the whole conversation, including every tool call and its result, returns a plan or a fix, and the executor gets back to work. The advisor never calls tools or writes output to you, it only advises.

It's the strategy Anthropic calls the advisor strategy: pairing Opus as counselor with Sonnet or Haiku as executor, to get near-Opus-level intelligence at Sonnet cost.

The numbers that matter

In Anthropic's tests, Sonnet with an Opus advisor gained 2.7 percentage points on SWE-bench Multilingual over Sonnet alone, while cutting cost per task by 11.9%. The most brutal case is Haiku: on the BrowseComp benchmark, Haiku with an Opus advisor scored 41.2%, more than double the 19.7% it gets alone.

How to turn it on

There are three ways: the /advisor command mid-session, which is saved as the default, the advisorModel setting in the settings file, or the --advisor flag at startup. Run /advisor opus and it's done. To turn it off, /advisor off.

The advisor has to be at least as capable as the primary model, so a Haiku can't advise a Sonnet. During the session you'll see an Advising line with the model name, and you press Ctrl+O to read the full advice.

The fine print

This is still experimental and not for everyone. You need Claude Code v2.1.98 or higher and it only works with the Anthropic API, so no Amazon Bedrock, Google Vertex AI or Microsoft Foundry. And it costs: each consultation spends tokens at the advisor model's rate, on top of what you already spend on the primary. Since the advisor only generates a short plan, typically 400 to 700 tokens, total cost stays well below running the strong model start to finish.

One detail developers will like: turning the advisor on or off mid-session does not invalidate the prompt cache on the primary model. Full documentation is here.