Agents & OrchestrationHow It Works

The advisor tool: Opus-level reasoning at Sonnet prices

In brief

A new Claude API feature lets Sonnet or Haiku call Opus mid-task when they need help. You pay Opus rates only for those calls — everything else runs at Sonnet or Haiku cost. Here's what it does and when to use it.

7 min read·AI Agent

Contents

♡Sign in to save

The advisor tool is a new feature in the Claude API (public beta, April 9 2026) that lets a lighter model — Sonnet or Haiku — call Opus when it needs help on a specific decision mid-task.

How it works: Sonnet runs the agent loop, handles tools, and works through steps the normal way. When it hits something it can't resolve — a hard reasoning problem, an ambiguous situation, a decision that needs more care — it calls the advisor. Anthropic runs a separate inference pass with Opus against the full conversation, Opus returns a plan or correction, and Sonnet continues.

This all happens within a single API call. No extra round trips from your side.

The cost logic is simple. Sonnet handles the bulk of the work. Opus only gets called when the executor decides it needs help — typically generating 400–700 tokens per call. Total cost stays well below running Opus end-to-end.

What the numbers show

Anthropic ran benchmarks and published the results at launch.

Sonnet 4.6 with Opus 4.6 as advisor:

SWE-bench Multilingual (software engineering tasks): +2.7 percentage points over Sonnet alone
Cost per agentic task: 11.9% less than Sonnet alone — the advisor helps Sonnet avoid wasted steps, so total task cost drops even accounting for the Opus tokens

Haiku 4.5 with Opus 4.6 as advisor:

BrowseComp (web research benchmark): 41.2%, up from 19.7% for Haiku alone
That is a 109% improvement on tasks that reward careful, multi-step reasoning

The Haiku result is the more striking one. Haiku alone has real limits on complex research tasks. With Opus advising it at the right moments, it more than doubles its score while still costing well below Sonnet or Opus running solo.

How to set it up

The advisor tool is declared like any other tool in the messages API. It requires the beta header.

import anthropic

client = anthropic.Anthropic()

response = client.beta.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=4096,
    betas=["advisor-tool-2026-03-01"],
    tools=[
        {
            "type": "advisor_20260301",
            "name": "advisor",
            "model": "claude-opus-4-6",
        },
        # your other tools here
    ],
    messages=[
        {"role": "user", "content": "Your task here"}
    ]
)

The executor decides when to call the advisor — you don't control that directly. Advisor tokens bill at Opus rates. Everything else bills at the executor model's rate.

Haiku as executor:

response = client.beta.messages.create(
    model="claude-haiku-4-5",
    max_tokens=4096,
    betas=["advisor-tool-2026-03-01"],
    tools=[
        {
            "type": "advisor_20260301",
            "name": "advisor",
            "model": "claude-opus-4-6",
        }
    ],
    messages=[...]
)

When to use it

The advisor tool fits agentic tasks — situations where the model is working through multiple steps, making decisions, and using tools. These are exactly the cases where judgment quality on specific sub-decisions matters most and where running Opus end-to-end becomes expensive.

Good fits:

Multi-step software engineering tasks (debugging sessions, PRs, refactors)
Web research and browsing workflows where reasoning quality varies by step
Any agentic loop where you want near-Opus quality on the hard decisions but not on every token

Less useful for:

Single-turn generation or Q&A — no multi-step loop, the advisor does not add much
Latency-sensitive applications — each advisor call adds roughly 1–3 seconds
Simple tasks Sonnet handles reliably on its own

The cost math

Running Opus end-to-end on a 10-step agentic task is expensive. Running Sonnet for all 10 steps is cheaper but may fail on the hard ones. The advisor pattern gives you a third option: Sonnet for the routine steps, Opus called in for the 1–2 decisions that actually need it.

In the SWE-bench results, Sonnet with Opus advisor cost 11.9% less per task than Sonnet alone. The efficiency gain on the overall loop offset the Opus token cost.

For Haiku with Opus advisor, you pay more than Haiku alone but still far less than Sonnet — while getting a quality jump that makes Haiku viable for tasks it previously could not handle.

Official docs

Full API reference: platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool

Announcement post: claude.com/blog/the-advisor-strategy