Anthropic releases advisor mode

My notes

Summary

Anthropic launched “Advisor mode” for Claude Code, where Opus acts as a planner/advisor while Sonnet or Haiku execute tasks — delivering better benchmark scores at lower cost. The key architectural insight is that Opus never makes tool calls, only provides plans and course corrections, which keeps token costs down while boosting agent quality.

Key Insight

• How it works: A cheaper model (Sonnet or Haiku) executes tasks while Opus serves as an advisor with full context. When the executing agent hits a problem, it escalates back to Opus for an updated plan.
• Opus does not make tool calls — this is the cost-saving mechanism. It only provides strategic guidance, not direct execution.
• Benchmark results: Sonnet with Opus Advisor scored 74.8 vs standalone Sonnet’s 72, at 0,96 USD vs 1,09 USD per task — better performance for ~12 % less cost.
• More sophisticated than simple plan-then-execute: The feedback loop where the agent can go back to the advisor mid-task distinguishes this from a basic “Opus plans, Sonnet codes” pipeline.
• Access: Available both via API and via /advisor command inside Claude Code.
• Pattern is generalizable: The advisor/executor split is a reusable architecture for any multi-model agent system where you want quality oversight without the cost of running the expensive model for every action.