Introduction

Anthropic has released a new API tool called Advisor Strategy, which allows Sonnet or Haiku to automatically consult Opus for guidance when encountering challenges during task execution. This reverse collaboration model, where smaller models work and larger models provide insights, brings the intelligence closer to Opus while keeping costs near Sonnet’s, potentially resulting in lower overall token consumption.

How It Works

The Advisor Strategy enables Sonnet (or Haiku) to act as the Executor, executing tasks, calling tools, reading results, and iterating. When it reaches a decision point where its judgment is insufficient, it consults Opus as the Advisor. Opus receives shared context and returns a plan, correction, or stop signal, allowing Sonnet to continue its work.

The Advisor does not call tools or produce user-facing outputs; it only provides guidance. Advanced reasoning only intervenes when the Executor needs it, and the entire process is billed at the Executor’s rate.

Evaluation Data

Sonnet + Opus Advisor

In the SWE-bench Multilingual tests, the combination of Sonnet and Advisor improved performance by 2.7 percentage points compared to Sonnet running solo, while reducing the cost per task by 11.9%. The cost reduction is attributed to the Advisor’s involvement, which helps the Executor avoid unnecessary detours and reduces total token consumption.

Haiku + Opus Advisor

In the BrowseComp tests, Haiku combined with Advisor scored 41.2%, more than double Haiku running solo (19.7%). Although this score is 29% lower than Sonnet’s solo performance, the cost was reduced by 85%.

Usage

Using the API is straightforward. Simply add an advisor_20260301 type tool to the tools array in the Messages API request, specifying the Advisor model as Opus and setting a max_uses limit for how many times the Advisor can be consulted per request.

The entire model handoff occurs in a single /v1/messages request, eliminating the need for additional network calls or manual context management. The Executor decides when to call the Advisor, and Anthropic routes the selected context to the Advisor model. After receiving a plan, the Executor continues execution.

Billing is based on the token usage of the Advisor and Executor models. The Advisor’s tokens are billed at Opus’s rate ($5/$25), while the Executor’s tokens are billed at Sonnet’s ($3/$15) or Haiku’s ($1/$5) rates. Since the Advisor typically generates a short plan (400-700 tokens), the overall cost is significantly lower than running Opus throughout.

Early User Feedback

Eric Simmons, CEO of Bolt, noted improved architectural decisions on complex tasks without extra overhead on simpler tasks. Kay Zhu, co-founder and CTO of Genspark, observed clear improvements in agent rounds, tool calls, and overall scores compared to their own planning tools. Anuraj Pandey, a machine learning engineer at Eve Legal, mentioned that Advisor allowed Haiku 4.5 to consult Opus 4.6 for structured document extraction, achieving state-of-the-art quality at a fraction of the cost.

Key Takeaways

1. This is the first time Anthropic has provided native support for model collaboration at the API level. Previously, coordinating Sonnet and Opus required custom orchestration logic and context management.
2. The pricing logic is clever; the Advisor outputs short plans (400-700 tokens) at a low cost, which can help the Executor avoid costly detours, leading to lower overall expenses.
3. The combination of Haiku and Opus Advisor is noteworthy, achieving competitive results at a significantly lower price, making it suitable for large-scale, cost-sensitive agent deployments.
4. Anthropic continues to accelerate its product offerings, with the recent releases of Mythos, Managed Agents, and the Advisor Tool, indicating rapid growth in their product line.