InferenceResearch

Your smartest model is doing your dumbest work.

Frontier agents spend ~12 steps searching before they edit. A disposable scout handles that at 1/9th the cost.

Frontier models shouldn’t be searching

Before a coding agent can fix a bug, it has to find the right files. That search takes ~12 steps on average — grepping, reading, navigating — and every one of those steps runs on the frontier model at ~$0.03 each. The model that should be reasoning about the fix is instead burning tokens on searching.

Delegate the search to a scout

In our harness a cheap, read-only scout agent handles the repository search instead. It runs on a small model, can’t edit anything, and gets thrown away after each delegation. When it’s done, it hands the planner a structured brief — byte-precise code snippets the frontier model needs to make informed edits. No wasted context, no searching.

What the paired runs show

The planner starts its work already looking at the right code. Across 50 paired tasks, delegating the search cut cost per solved task on every model we ran — by a fifth to a half, depending on the model. That is what delegation buys, and on this benchmark it is the whole of what it buys.

Accuracy is a separate question and these runs do not answer it. The resolve rate moved up on two of the four models and down on the other two, every move well inside the spread our own seed-only re-draw produces from sampling alone; and our gold-patch coverage result found that failures here are implementation rather than retrieval, which bounds how much better search could buy in the first place. Both of those are our own results, and they are why the accuracy column below is reported rather than claimed.

The harness itself reproduces Anthropic’s published SWE-bench Pro figure for the same model, which is what licenses trusting the numbers on this page — a match, and not a win.

Benchmarks

Frontier models waste money searching.

Each model ran 50 real GitHub issues from SWE-bench Pro — once normally, and once with a scout handling the search.

SystemnBaselineWith Scout$/solvedΔ costΔ accuracy
Claude Opus 4.85033/5066%36/5072%$1.23was $1.75−29.8%+3
Claude Opus 55034/5068%37/5074%$1.27was $1.61−21.0%+3
GPT-5.6 Sol5036/5072%34/5068%$0.82was $1.05−21.9%−2
GLM-5.25028/5056%27/5054%$0.60was $1.25−52.0%−1

SWE-bench Pro · 33/50 = 33 of 50 issues solved · $/solved = total spend ÷ solved count · Δ accuracy is reported, not claimed — every delta here sits inside the spread our own seed-only re-draw produces from sampling alone · Naïve measured · July 2026

Step economics

Move discovery to the cheaper model.

One planner step costs about what nine scout delegations cost.

One planner step$0.0311

expensive model · full agent context

One scout delegation$0.0034

cheap model · ~7 read-only steps · 9× cheaper

Without a scout, the planner spends ~12 steps searching the repo before it can edit, every one of them on the expensive model. The scout handles that discovery at 1/9th the cost.

ModelBaseline search stepsWith scoutReduction
Sol (GPT-5.6)11.6 steps6.4 steps−44%
Opus 511.6 steps8.1 steps−30%
Search steps per task · n=50 paired · each step costs ~$0.03 on the planner
About this note
Status
Research
Focus
Scout discovery before planner judgment
Surface
Early / evolving
Papers
Findings land on /lab/papers when gated

Follow the work

Early Lab. Papers and findings land here when they clear the gate.