The advisor strategy: Give Sonnet an intelligence boost with Opus | Claude

• The advisor strategy pairs Sonnet/Haiku as executor with Opus as advisor—Sonnet runs the task end-to-end and only consults Opus when stuck, inverting the typical orchestrator/worker pattern
• Sonnet+Opus advisor scored 2.7pp higher on SWE-bench Multilingual while costing 11.9% less per task than Sonnet alone
• Haiku+Opus advisor scored 41.2% on BrowseComp (vs 19.7% solo), trailing Sonnet by 29% but costing 85% less—ideal for high-volume tasks
• Implementation is a one-line tool declaration in the API—no extra round-trips, context management, or orchestration logic required
• Built-in cost controls via max_uses parameter, with advisor tokens billed separately so you can track spend per tier

The advisor strategy fundamentally inverts how most developers architect multi-model AI systems. Instead of using a powerful model as an orchestrator that decomposes work and delegates to cheaper worker models, you use the cheap model (Sonnet or Haiku) as the primary executor that drives the entire task—calling tools, reading results, iterating toward solutions. Only when the executor hits a decision it can't solve does it escalate to Opus for guidance. Opus never calls tools or produces user-facing output; it only provides plans, corrections, or stop signals based on the shared context, then hands control back to the executor. This means frontier-level reasoning applies only when needed, keeping most of the run at executor-level cost.

The benchmarks validate the approach across multiple dimensions. Sonnet with an Opus advisor improved 2.7 percentage points on SWE-bench Multilingual over Sonnet alone while reducing cost per task by 11.9%. On BrowseComp and Terminal-Bench 2.0, Sonnet+advisor outperformed Sonnet solo while costing less per task. The pattern also works down the model tier stack: Haiku with an Opus advisor scored 41.2% on BrowseComp (more than double its 19.7% solo score), trailing Sonnet solo by only 29% in accuracy but costing 85% less per task—making it viable for high-volume workloads that need intelligence but can't afford Sonnet at scale.

Claude implemented this as a server-side tool that requires zero orchestration logic from developers. You declare advisor_20260301 in your Messages API request with a model parameter (typically claude-opus-4-6) and optional max_uses cap, and the entire handoff happens inside a single API call. The executor decides when to invoke it, Claude routes the curated context to the advisor model, returns the plan, and the executor continues—all with no extra round-trips or context management. Advisor tokens are billed at advisor rates, executor tokens at executor rates, and both are reported separately in the usage block so you can track spend per tier. The advisor tool works alongside your existing tools, so your agent can search the web, execute code, and consult Opus in the same loop.