/p/2026-10-15 · explainer
Paper explainer · 2610.15093 · Kumar and Pandey

You trimmed the logs. The cost was in the turns.

A coding agent re-sends its whole conversation every turn, so the frightening line item is input tokens — except that 98.6% of them were served from the provider’s prefix cache, which makes a flat input price overstate the real bill about sevenfold. Fit cost against what actually varies across 600 trajectories and turn count explains 84% of it; the average size of what the tools printed back explains 4%. The lever is the loop, not the logs.

01 · The problem

The number on the dashboard is not the bill

the shape of one trajectory

02 · The mechanism

Halve every log and the bill barely notices

03 · The evidence

Turn count, 84 percent. Observation size, 4.

which model, in which scaffold

04 · Why I care

Delegation has a break-even, and cached reads move it

which price you measure the saving against

how big an observation has to be before the policy touches it

05 · Apply it

Your loop, your prices illustrative

Results

What the paper actually measured

What it does not show

In practice