Anthropic Just Made Cheap Models Think Like Opus (Claude Advisor Tool Is Wild)

Source: Two articles in the Email 08 bundle (2026-04-13):

  • Anthropic Just Made Cheap Models Think Like Opus (Claude Advisor Tool Is Wild)
  • Claude’s Advisor Tool: The Sub-Agent Pattern, Inverted

Key takeaways

  • The sub-agent pattern, inverted. Normally: Opus delegates to cheaper sub-agents. Advisor: cheaper models consult Opus for difficult moments.
  • “Makes cheap models think like Opus” — runs most of the session on Haiku/Sonnet but pulls in Opus’s reasoning when the current model gets stuck.
  • Direct implication for cost-routing architecture: you can ship most workloads on cheaper models and still get Opus-level outcomes on hard problems.

The inversion explained 

Normal pattern:
  Opus (expensive, always available)
    └─ delegates to Sonnet/Haiku for simple subtasks

Advisor pattern:
  Sonnet/Haiku (cheap, default)
    └─ pulls in Opus *only* when stuck

The normal pattern starts expensive and tries to save money by delegating down. The Advisor pattern starts cheap and consults up. Cheaper in the average case, still gets Opus quality when it actually matters.

Why this matters for cost-routing

For any Harris / Constellation product that routes between model tiers:

  • Current architecture often: pick a tier, stick with it for the session
  • Advisor pattern alternative: start Sonnet, escalate to Opus for specific hard moments
  • Combined with task budgets (Opus 4.7 public beta): reasoning-depth becomes a per-task dial

This could materially change ACT’s model-routing logic. Worth reading both articles to understand the mechanism.

Built by Force Information Systems · Harris Computer · Constellation Software. Licensed under MIT.