Sonnet Does the Work, Opus Reviews: Does Claude Code's advisor Save Money?

Anthropic has added an experimental advisor feature to Claude Code. Claude Code is Anthropic's command-line coding assistant, and the new twist is dividing labor between models: Sonnet handles everyday work, while Opus reviews at key moments. Both are Anthropic models; Sonnet is lightweight and cheap, while Opus is the most capable but expensive.

How to use it

At its core, it's just one command:

claude --model sonnet --advisor opus

Sonnet runs the entire session and is responsible for writing code. Opus does not act directly; instead, it reads the full conversation transcript (including every tool call and result) and then gives directional advice. According to the official docs, Claude tends to consult the advisor at three points: before settling on an approach, when errors keep recurring, and before wrapping up a task.

There are three ways to set it up:

Add the --advisor opus argument at startup
Enter /advisor opus during a session to switch
In the config file, set "advisorModel": "opus"

You can also ask directly in the prompt: "consult the advisor before you continue".

Actual interface

This is the interface during actual operation. On the left, the Opus 5.5 advisor is called in to review the plan; in the middle is the Sonnet 5.5 execution layer; on the right are Haiku 5.5 subagents for search, grep, and auditing. In a single task, three layers of models collaborate, each doing what it is best at.

What the community is arguing about

Cost savings are the core controversy. Supporters say this is like having a senior engineer on a team: you don't need the big shot for routine small matters—call them in only when you're stuck. Opponents say every advisor call is billed at Opus rates, so if Claude calls too often, the extra calls will eat up the savings. Some users have named this expense: advisor tax.

Limitations

Experimental feature; behavior may change
Anthropic API only; Bedrock and Vertex are not supported
Advisor calls are billed at Opus rates

A brief take

My take is that the real value of this feature is turning "have a strong model step in only at key points" into a product capability. Previously, achieving the same effect required building your own workflow; now it takes one command.

But giving Claude itself the decision-making power over spending is a subtle design choice. The official wording is "tends to consult", not "consults only when necessary". Whether it saves money will only be clear after running a few real tasks and checking the bill with /usage.

For full details, see the official documentation: https://code.claude.com/docs/en/advisor