Stencil ran the “senior architect, junior engineer” pitch to ground and it lost. In their SWE-Bench Pro numbers, Opus planning read-only while a cheap model executes lands at $3.18 per task; Opus doing the whole thing alone is $2.78. The “cost-saving” handoff costs 14% more.

• 💡 The bill is O(reads), not O(edits). Across 1.81B tokens, actually writing files was ~9% of spend. Reading — files, dead ends, hypotheses tested — is what scales the cost, and every model pays full price for it.
• ⚠️ A plan doc is a 2K-token postcard distilled from 100K+ tokens of grounded context. The executor gets the postcard, not the understanding, then rebuilds the rest at its own price.
• 🔁 /plan duplicates the expensive part. The frontier model reads everything at frontier prices; the cheap model re-reads the same files at its prices. You didn’t move the cost — you paid it twice.
• 🎯 Hand off the trajectory, not the fairytale. Their /prewalk keeps the frontier model running until the first edit lands — the moment it’s confident enough to act — then swaps to the cheap model and prunes the planning instruction from context.
• 📉 The cheap model never notices the swap. It sees an explored plan, a todo list, and one valid move already made — a free in-context example — and just keeps going. Result: $1.46 per task at ~97% of frontier quality.

This matches what I see profiling agent cost: teams optimize the wrong turn. They cap the executor and leave the reader untouched. The HN discussion has people noting the split can still pay off when planner and executor share a warm cache — which is the same point from the other side: transfer context, not prose.

If reading is the real bill, why is every agent design still priced like the thinking is the expensive part?