Anthropic has officially launched the Advisor Tool in the Claude Platform API, offering developers a new hybrid intelligence model: a cheaper and faster model handles tasks throughout, while automatically consulting the most powerful model for strategic advice when facing complex decisions. This design completely overturns the traditional "large model breaks down tasks, small model executes" Agent development approach, achieving a true "cost-performance optimal solution."
Innovative Mechanism: Small Model Executes, Large Model Acts as Advisor
In this new feature, Claude Sonnet or Haiku takes on the "executor" role, responsible for tool calls, task processing, and result integration. Only when the executor faces a difficult problem it cannot resolve, will it pass the current context to Claude Opus as an advisor.
Opus provides high-level solutions, corrects direction, or offers strategic guidance, then immediately returns to the executor to continue. Throughout this process, Opus does not directly call tools or output final results to users, acting as an invisible "military strategist." This reverse design ensures that most Token consumption occurs on the lower-cost model, using high-end computing power only at critical points.
Performance and Cost Win-Win Test Results
According to official test data:
Sonnet + Opus Advisor improved by 2.7 percentage points on the multi-language version of SWE-bench compared to Sonnet alone, while the overall cost per task decreased by 11.9%.
Haiku + Opus Advisor performed even more impressively, with its score rising from 19.7% to 41.2% in the BrowseComp test, doubling its performance. Although it still lags behind Sonnet's single model performance by about 29%, its cost is only 15% of Sonnet's, making it ideal for high-concurrency, high-tolerance batch scenarios.
This achievement fully demonstrates that through intelligent routing of key decisions, developers no longer need to make a tough choice between "using Opus all the time is too expensive" and "using only Sonnet is not strong enough."
Simple Usage, Enable with One Line of Configuration
Developers just need to add a tool definition of type advisor_20260301 to the tools array in the Messages API. Complex operations like model switching and context passing are automatically handled within a single API request, without the need for manual management of multiple calls or additional context.
You can also flexibly control the maximum number of times the advisor is consulted per request using the max_uses parameter. In terms of billing, the Token consumption of the executor and the advisor will be billed separately, ensuring transparency and control. Currently, this feature is in the Beta phase, and you need to add anthroic-beta: advisor-tool-2026-03-01 to the request header to use it.
Industry Significance: Opening New Cost-Performance Doors for Agent Development
AIbase believes that the launch of the Advisor Tool marks an important evolution in AI Agent development paradigms. It allows developers to enjoy the execution efficiency of Sonnet for 95% of the time at nearly Sonnet's price, while gaining Opus-level judgment at 5% of the key decision points. This "executor + advisor" hybrid strategy not only significantly reduces the cost of large-scale deployment but also greatly improves the usability of small models in complex tasks.
Whether building a code agent, browser automation, or high-volume data processing scenarios, the Advisor Tool provides an efficient, flexible, and economical option. Anthropic once again proves through action that breakthroughs in model capabilities come not only from parameter scale, but also from innovations in architecture and usage patterns.
As the Beta phase progresses, this feature is expected to quickly become a standard tool for developers building cost-effective AI Agents, and it is worth closely watching its subsequent updates and more benchmark performances.
