Anthropic · 2026-04-09 · major
Anthropic Advisor Tool — Pair Sonnet or Haiku with an Opus Advisor
Claude API beta: a cheap executor (Sonnet or Haiku) handles tasks end-to-end and calls Opus 4.6 for strategic guidance when needed. Haiku with Opus advisor scored 41.2% on BrowseComp vs 19.7% solo, at 85% lower cost than Sonnet.
A Haiku or Sonnet executor calls Claude Opus 4.6 mid-task for strategic planning — one API call, near-Opus intelligence at executor rates.
Key specs
| Haiku browse comp solo | 19.7% |
|---|---|
| Haiku browse comp with opus advisor | 41.2% |
| Sonnet swe bench multilingual gain | +2.7 pp |
| Cost vs sonnet alone | -11.9% |
What is it?
The advisor tool is a beta Claude API feature that lets a small executor model (Sonnet 4.6 or Haiku 4.5) pause and consult Claude Opus 4.6 for strategic advice at key points in a long task. The advisor reads the full conversation history and returns a structured plan; the executor then continues. Everything happens inside a single API request with no extra client-side round trips.
How does it work?
Add {"type":"advisor_20260301","name":"advisor","model":"claude-opus-4-6"} to the tools array along with the beta header anthropic-beta: advisor-tool-2026-03-01. When the executor encounters a hard decision, it emits a server_tool_use block with name "advisor". Anthropic runs a separate Opus inference pass server-side — the advisor sees the full transcript and returns advice as an advisor_tool_result block. A max_uses parameter caps advisor calls per request, and an optional caching field enables prompt caching for long multi-turn loops.
Why does it matter?
Most agent tasks are mechanically simple with a few genuinely hard decision points. Paying Opus rates for every token wastes money; using Haiku throughout leaves intelligence on the table at the hard moments. The advisor pattern applies Opus-level planning only where it is needed. In early benchmarks, Haiku with Opus advisor scored 41.2% on BrowseComp vs 19.7% solo — more than double — while costing 85% less per task than switching the executor to Sonnet.
Who is it for?
Claude API users building coding agents, computer-use pipelines, and multi-step research workflows.
Try it
anthropic-beta: advisor-tool-2026-03-01 + type: advisor_20260301 in tools array