Article analysis
Cost per successful task → measures → Per-request reasoning_effort levels
The author proposes cost per successful task as the basis for evaluating reasoning levels. A cost advantage requires lower total consumption under the same success criteria; the availability of adjustable levels alone does not demonstrate savings.
Read the evidence
Inference evaluation should record the selected level, actual input length, number of attempts, final success rate, total token count, and human review time. Only when a level reduces total consumption per successful task under the same success criteria can it be judged to lower costs for that workload.
Background sources for the article's reasoning: [S5] Alibaba Cloud
Read the full research