Research

Papers and experiments

• White paper · Model routing in agentic coding

  Cheapest per token is not cheapest per task

  A pre-registered experiment on 780 graded agent sessions. Letting the parent pick the worker model cut cost per completed task 76% against always-Opus, but a fixed always-Sonnet default beat the router by another 23%, and the cheapest model per token cost 2.4× Sonnet per task.

  Read Cheapest per token is not cheapest per taskSeptember 2026 · 16-page PDF

• Hypothesis paper · Governing AI model spend

  Draft · not yet tested

  Route on evidence, not on the prompt

  A proposed escalation ladder: start every session on a mid-tier model, keep frontier intelligence one evidence-backed step away, and decide the model on what happens during the work rather than on the prompt. Includes the protocol for testing it against the best fixed default.

  Read Route on evidence, not on the promptSeptember 2026 · 19-page PDF