Anthropic launched the Advisor Strategy on April 9, 2026 — a way to get near-Opus intelligence in your AI agents at a fraction of the cost.
The idea is simple. Instead of running Opus on every request, you pair a cheaper model with Opus as an on-demand advisor. The cheap model does all the work. Opus only steps in when needed.
How It Works
You have two roles:
- Executor — Sonnet or Haiku. Runs every turn. Calls tools, reads results, makes decisions.
- Advisor — Opus. Runs on-demand only. Reviews the shared context and sends guidance when the executor is stuck.
The executor and advisor share the same context: the system prompt, tool definitions, full conversation history, and all prior tool results. When the executor hits a hard decision, it calls the advisor tool. Opus reviews the full context and sends back a plan or correction. Then the executor continues.
The advisor never calls tools directly. It only advises.
The Numbers
The performance gains are significant:
| Setup | Result |
|---|---|
| Haiku alone (BrowseComp) | 19.7% |
| Haiku + Opus advisor (BrowseComp) | 41.2% — more than 2× |
| Sonnet + Opus advisor (SWE-bench) | +2.7 percentage points |
| Cost vs Sonnet solo | 85% cheaper (Haiku + Opus advisor) |
| Cost per task (Sonnet + Opus advisor) | 11.9% reduction |
In Anthropic’s evals, Haiku with an Opus advisor more than doubled Haiku’s solo score on BrowseComp, while Sonnet with an Opus advisor improved SWE-bench results at lower per-task cost than Sonnet alone. The advisor only generates around 400–700 tokens per consultation, so the Opus cost stays low.
How to Use It
The feature is in beta. Add the beta header and declare the advisor tool in your Messages API request:
POST /v1/messages
anthropic-beta: advisor-tool-2026-03-01
{
"model": "claude-sonnet-4-6",
"tools": [
{
"type": "advisor_20260301",
"name": "advisor",
"model": "claude-opus-4-6",
"max_uses": 3
}
],
"messages": [...]
}
The model handoff happens inside a single /v1/messages request. No extra round trips. No context management. No separate API calls.
max_uses controls how many times per request the executor can call the advisor. Anthropic bills advisor tokens separately at Opus rates — but since the advisor only runs when needed and generates short responses, the total cost stays low.
When to Use It
The Advisor Strategy works best for long-running agentic tasks where:
- Most steps are straightforward (executor handles them)
- A few steps need deep reasoning (advisor steps in)
- Cost matters (you can’t run Opus on every request)
Good examples: code agents, browser automation, multi-step research tasks, data pipelines.
It is less useful for simple, single-turn requests where you just want a fast answer.
Why This Matters
Most teams face a hard choice: use a powerful model and pay more, or use a cheap model and accept lower quality. The Advisor Strategy breaks that trade-off.
You get the speed and cost of Haiku for routine work. You get Opus reasoning for the hard parts. And you pay for Opus only when you actually need it.
Read the full technical breakdown on the Anthropic blog.
Related Articles