Anthropic is formalizing a two-model workflow pattern on the Claude Platform: use Opus as the "advisor" - the model that plans, reasons, and decides - while Sonnet or Haiku handles the actual execution.
The idea is straightforward. Opus is Anthropic's most capable model but also the most expensive and slowest. Sonnet and Haiku are cheaper and faster but less capable at complex reasoning. The advisor strategy threads the needle: Opus figures out what to do, then a lighter model does the work. For tasks that involve many steps - writing a report, processing a batch of documents, running a multi-step research loop - this can cut costs substantially without sacrificing quality on the decisions that matter most.
How It Works in Practice
In this pattern, Opus receives the full task context and produces a plan or set of instructions. Sonnet or Haiku then executes against those instructions, only looping back to Opus when something unexpected happens or a judgment call is needed. You're paying Opus rates for the high-stakes reasoning, Haiku rates for the routine execution.
This is not a new concept - developers have been building similar patterns manually for months. What Anthropic is doing now is making it a first-class supported pattern on the platform, which means better documentation, cleaner API patterns, and presumably some tooling to make the handoff between models smoother.
Cost Math for API Users
For anyone building Claude-powered products, the cost difference between running everything through Opus versus a hybrid Opus-Haiku setup is significant. Opus runs at $15 per million input tokens and $75 per million output tokens. Haiku is $0.80 and $4 respectively. If 80% of your token usage is execution-level work that Haiku can handle, the savings are real.
The advisor strategy also has implications for response quality. Cheaper models can drift or miss edge cases on complex, multi-step tasks. Keeping Opus in the loop as a supervisor - rather than handing everything to Haiku and hoping - is a practical tradeoff that many teams have already landed on independently.
The announcement signals Anthropic is paying attention to how production teams actually use their models, not just how researchers benchmark them.