Advisor Strategy: Boosting Sonnet's Intelligence by Using Opus as an Advisor

The Advisor strategy has been officially introduced to the Claude Platform. This strategy is a pattern that combines the high-performance model Opus as an advisor with Sonnet or Haiku as the executor, achieving high reasoning capability in a cost-efficient way.

Key Features and Performance Results

• Performance improvement and cost reduction: When Sonnet is combined with an Opus advisor, the SWE-bench Multilingual score improved by 2.7 percentage points, while cost per task decreased by 11.9%. When Haiku is combined with Opus, the BrowseComp score reaches 41.2%, more than double that of Haiku alone (19.7%), while cutting costs by 85% compared to Sonnet alone.
• Operating mechanism: When the executor reaches a decision it struggles to resolve, it requests guidance from Opus. Opus does not directly call tools or generate output; instead, based on the shared context, it provides only guidance such as plans, corrections, and stop signals.
• API implementation: Declaring the advisor_20260301 tool in a Messages API request completes the model handoff within a single request. It works without separate context management or additional round-trip requests, and the number of calls can be limited via the max_uses parameter.
• Cost structure: Advisor tokens are billed at Opus rates, while executor tokens are billed separately at the executor model's rates. Since the advisor generates only short plans (approximately 400-700 tokens), overall costs can be kept low.