Anthropic just dropped what they’re calling the “advisor strategy,” and it’s basically Claude telling us: stop running everything through the best (and most expensive) model. The idea is simple: you pair Opus as the advisor with a cheaper executor model like Sonnet or Haiku. The executor does the work, and it only calls Opus when the task actually needs that higher-level reasoning. That’s how you get near-Opus intelligence in your agents without paying Opus prices for every single step. In this video, I break down the model pricing (output tokens are the real budget killer), explain the difference between the Messages API (where advisor strategy exists) vs Claude Code (the terminal product), and then I demo a dashboard I built to test easy, medium, and hard prompts across different setups: Haiku+Opus, Sonnet+Opus, Sonnet solo, and Opus solo. The big takeaway: don’t blindly ship this—test hundreds of prompts for your use case. Then I show you how to apply the same “use Opus only when needed” mindset inside Claude Code using the hidden “opus plan” mode so your session limit lasts way longer while still getting strong plans and clean execution.