Count what the observer processed.
The operator’s monthly exports total about 125.1 million generation tokens. His estimate for observation generation was $180–$250 using the model mix and pricing available during the study. These tokens are processing volume, not tokens saved.
Budget for the work you will actually run.
The author’s wider $225–$295 estimate also included about $44 for a priced subset of model-benchmark runs. Your observation provider, capture volume and pricing determine your costs. The study estimate covers this one deployment in spring 2026.