Anthropic Advisor Strategy: Sonnet Runs, Opus Guides – Scores Up, Costs Down

Advisor Strategy Overview

In the Advisor Strategy architecture the cheaper model (Sonnet or Haiku) runs the entire task as the executor . When the executor reaches a decision point it cannot resolve, it sends the current context to Opus, which acts as an advisor . Opus returns a short plan (typically 400‑700 tokens) or a signal to stop, without invoking any tools or producing user‑facing output. The executor then continues using the advice.

This inverts the traditional “large model orchestrates, small model executes” pattern: the small model works end‑to‑end and only consults the larger model when needed, eliminating the need for separate orchestration logic or a worker pool.

Benchmark Results

SWE‑bench Multilingual : Sonnet + Opus Advisor improves accuracy by 2.7 percentage points over Sonnet alone while reducing per‑task cost by 11.9 %.

BrowseComp : Haiku alone achieves 19.7 % success; Haiku + Opus Advisor reaches 41.2 % (more than double) at only 15 % of Sonnet’s cost.

Terminal‑Bench 2.0 : similar gains are observed (exact numbers omitted in source).

The cost reduction stems from Opus generating only a brief plan instead of the repeated tool calls and token consumption that Sonnet would incur when stuck.

Integration Guide

Add a single advisor entry to the tools array of the Messages API. Example:

response = client.messages.create(
    model="claude-sonnet-4-6",  # executor
    tools=[
        {
            "type": "advisor_20260301",
            "name": "advisor",
            "model": "claude-opus-4-6",
            "max_uses": 3,
        },
        # ... your own tools
    ],
    messages=[...]
)

Key Parameters

Billing separation : tokens used by the advisor are billed at Opus rates; executor tokens are billed at Sonnet/Haiku rates.

max_uses limits how many times the advisor can be invoked per request, providing a hard cap on cost.

Single‑request loop : the entire advisor interaction occurs within one /v1/messages call; no extra round‑trips are required.

Tool compatibility : the advisor tool can be listed alongside other tools such as web search or code execution.

The beta feature requires the request header anthropic-beta: advisor-tool-2026-03-01.

Concrete Use Case

An engineer reported that on a structured‑document extraction task, the combination Haiku 4.5 + Opus 4.6 Advisor achieved flagship‑model quality while costing only one‑fifth of the baseline.

References

Anthropic blog – Advisor Strategy: https://claude.com/blog/the-advisor-strategy

Claude Platform Advisor Tool documentation: https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool