Achieve Top‑Tier AI Performance at Low Cost with Claude’s Advisor Strategy

Claude has introduced an "Advisor Strategy" that enables a low‑cost model (e.g., Sonnet or Haiku) to act as the executor while consulting a top‑tier model (Opus) for guidance only when it encounters a difficult sub‑task. The executor runs the end‑to‑end task, calls tools, reads results, and iterates solutions; the advisor provides a plan, a correction, or a stop signal and never directly generates user‑facing output.

Performance impact. In Claude’s own evaluation, pairing Sonnet with an Opus advisor on the SWE‑bench Multilingual benchmark raised the score by 2.7 percentage points compared with Sonnet alone, while overall task cost dropped by 11.9%.

For the lighter Haiku model, the effect is even more dramatic. On the BrowseComp benchmark, Haiku with an Opus advisor achieved a 41.2% score—more than double its solo score of 19.7%—and each task’s cost fell by 85%, making it ideal for high‑throughput scenarios that need a balance of intelligence and expense.

One‑click adoption via the new "advisor tool". Claude’s API now includes a server‑side tool named advisor_20260301. Developers simply declare this tool in a Messages API request; the model switch happens within a single request without extra network round‑trips or complex context handling. The executor decides when to invoke the advisor.

Example request:

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

Billing and cost control. Tokens generated by the advisor are billed at Opus’s rate, while tokens from the executor are billed at the cheaper Sonnet/Haiku rate. Because the advisor typically emits only short plans (≈400‑700 tokens), total cost is far lower than using Opus for the entire task. The max_uses parameter further caps advisor usage per request.

Developer feedback. Eric Simmons, CEO of Bolt, notes that the advisor makes better architectural decisions on complex tasks without adding overhead on simple ones. Anuraj Pandey, ML engineer at Eve Legal, reports that on structured document extraction, Haiku 4.5 with the advisor matches cutting‑edge model quality while cutting cost five‑fold.

Getting started. The advisor tool is available in beta. To enable it:

Add the beta header anthropic-beta: advisor-tool-2026-03-01 to your API request.

Include the advisor_20260301 tool in your Messages API payload.

Adjust system prompts to suit your scenario.

Claude recommends running existing evaluation suites to compare three setups—Sonnet alone, Sonnet with Opus advisor, and Opus alone—to find the optimal trade‑off for your use case.