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

• The advisor strategy pairs Sonnet/Haiku as executor with Opus as advisor—the cheap model does all the work and only consults the expensive model for guidance on hard decisions
• Sonnet+Opus advisor shows 2.7pp improvement on SWE-bench while reducing cost per task by 11.9%; Haiku+Opus doubles performance at 85% lower cost than Sonnet alone
• Implemented as a single tool declaration in the API—no extra round-trips, context management, or orchestration logic needed
• Advisor only generates short plans (400-700 tokens) while executor handles full output at its lower rate, keeping overall cost well below running Opus end-to-end
• Inverts the common pattern where a large orchestrator decomposes work for smaller workers—here the worker escalates without decomposition

Anthropic's advisor strategy fundamentally rethinks AI agent architecture by inverting the typical orchestration pattern. Instead of having an expensive model (Opus) orchestrate and delegate to cheaper workers (Sonnet/Haiku), a cheaper model drives the entire execution and only escalates to the expensive model when it hits a decision it can't solve. Sonnet or Haiku runs the task end-to-end—calling tools, reading results, iterating—and only consults Opus for guidance when stuck. The advisor never calls tools or produces user-facing output; it only provides plans, corrections, or stop signals before the executor resumes.

The benchmarks show this isn't just theoretically elegant—it delivers measurable improvements. Sonnet with an Opus advisor improved 2.7 percentage points on SWE-bench Multilingual while reducing cost per task by 11.9%. The pattern works even better with Haiku as executor: Haiku+Opus scored 41.2% on BrowseComp (more than double Haiku's solo 19.7%), trailing Sonnet solo by only 29% in score but costing 85% less per task. The cost efficiency comes from the advisor only generating short plans (typically 400-700 tokens) while the executor handles the full output at its lower rate.

The implementation is remarkably simple—declare advisor_20260301 as a tool in your Messages API request, and the model handoff happens server-side within a single request. No extra round-trips, no context management, no orchestration logic. The executor decides when to invoke it, Anthropic routes the curated context to the advisor, returns the plan, and the executor continues. You can set max_uses to cap advisor calls per request, and advisor tokens are reported separately 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.