Where to Spend the Reasoning Model

Your deep research pipeline has a fixed compute budget and someone proposes: "let's run every stage on the reasoning model — it's a high-stakes report, we shouldn't cheap out." Push back on this using the specific stage-by-stage shapes in this pipeline.

1. Name two stages where a reasoning model earns essentially nothing over a cheap model, and explain the task-shape reason why.
2. Name the one stage you would defend spending the reasoning model on even under real budget pressure, and justify it in terms of error cost, not just task difficulty.
3. A sub-agent's own analysis step is sometimes worth escalating to a reasoning model and sometimes not, even though every sub-agent runs the "same" step. What's the concrete signal that should decide it per sub-agent, rather than deciding it once for the whole stage?