Anthropic's Advisor Strategy: Opus-Level Intelligence at Sonnet Prices

Anthropic released a new API pattern called the advisor strategy that pairs Opus as a planning advisor with Sonnet or Haiku as the executor. The result is better benchmark scores than Sonnet alone, at a lower cost per task. If you are building applications on Anthropic's API outside of Claude Code, this is worth implementing immediately.

How the Advisor Strategy Works

The pattern is straightforward. Opus serves as the advisor and Sonnet (or Haiku) serves as the executor. Sonnet runs the main loop, makes all the tool calls, and does the actual work. Opus sits alongside it with full shared context, ready to provide guidance when the executor hits a decision it cannot reasonably solve on its own.

This is similar to how Claude Code handles Opus in plan mode with Sonnet executing, but with one important difference: the back-and-forth is continuous, not one-shot. In Claude Code's plan mode, Opus gives a strategy once and Sonnet runs with it. With the advisor strategy, Sonnet can consult Opus multiple times throughout execution whenever it encounters a stumbling block. The relationship stays active for the entire task.

Why Opus Never Makes Tool Calls

To keep costs low, Opus never makes tool calls. All tool execution stays with the cheaper model. Opus only reads the shared context and sends advice back to the executor. This design means you get Opus-level reasoning on the hard decisions without paying Opus prices for every API interaction in the chain.

The Benchmark Numbers

The performance data from Anthropic's blog post backs up the value claim. Comparing Sonnet 4.6 (medium effort) solo against Sonnet 4.6 (medium effort) with an Opus advisor:

SWE-Bench: 74.8% with advisor vs. 72.1% solo
BrowseComp (agentic search): 60.4% with advisor vs. 58.1% solo, and the cost dropped from $7.00 to $6.13
Terminal-Bench 2.0 (agentic coding): 63.4% with advisor vs. 59.6% solo, cost dropped from $0.94 to $0.88
Cost per agentic task on SWE-Bench: $0.96 with advisor vs. $1.09 solo

The pattern is consistent across benchmarks. Scores go up and costs go down.

How to Implement It

This is an API-level feature, not a Claude Code feature. To use it, you adjust how your code makes API calls. Specifically, you set the type to advisor and configure the max_uses parameter, which controls how many times the executor can consult Opus during a single task. That parameter lets you tune the cost-performance tradeoff for your use case.

Who Should Use This

The advisor strategy fills a gap that Anthropic users have felt for a while. Opus is overkill for most tasks, but Sonnet sometimes falls short on complex reasoning. There was no middle ground. Now there is one, and it costs less than running Sonnet alone. Anyone running Anthropic API calls in production web applications, agent frameworks, or automated pipelines should evaluate this pattern.

Key Takeaways

Opus advises, Sonnet executes. Opus provides reasoning on hard decisions while Sonnet handles all tool calls and routine work.
The relationship is continuous. The executor can consult the advisor multiple times per task, unlike Claude Code's one-shot plan mode.
Benchmarks show higher scores at lower cost. SWE-Bench, BrowseComp, and Terminal-Bench all improved while cost per task decreased.
Implementation requires two API parameters. Set the type to advisor and configure max_uses to control how often the executor consults Opus.
Best suited for API-based applications. If you are already inside Claude Code, this pattern is handled differently. The advisor strategy targets custom applications built on Anthropic's API.