Descriptions:
Anthropic recently introduced an adviser strategy that lets developers pair the high-capability Claude Opus model as an on-demand adviser with a cheaper executor model — either Sonnet or Haiku — so that expensive reasoning is only invoked when the task genuinely requires it. Nate Herk’s video unpacks how the strategy works, when to use it, and how to implement it today in both the Messages API and Claude Code.
The pricing math is central to the pitch: Opus costs $5 per million input tokens and $25 per million output tokens, compared to Sonnet at $3/$15 and Haiku at $1/$5. In Anthropic’s own evaluations, Sonnet paired with Opus as adviser scored 2.7 percentage points higher on SWE-bench than Sonnet alone and reduced cost per agentic task by nearly 12%. Using Haiku as the executor with Opus advising pushed browse comp scores from 19.7% (Haiku solo) to 41.2% — more than doubling performance at a cost still below running Opus alone. A live cost comparison on a support-bot task showed Opus running 21 times more expensive than Haiku for the same correct answer.
Herk also demonstrates the strategy on a customer-support chatbot, showing easy queries handled by Haiku without escalation and a complex multi-condition return-policy question triggering the adviser. He explains the architectural difference between the Messages API (where the adviser function lives as an HTTP endpoint) and Claude Code (a finished product built on top), and provides a GitHub repo for viewers to test the setup themselves.
📺 Source: Nate Herk | AI Automation · Published April 09, 2026
🏷️ Format: Tutorial Demo
