Flat per-run pricing for AI work is wrong in both directions. A one-line lookup subsidises a twenty-page crawl, and a run that failed to find anything still costs you the same as one that succeeded.
AI enrichment now charges on measured usage: the tokens the model consumed and the bytes the agent fetched. Two consequences worth knowing:
- Cheap questions are cheap. A short prompt against a small page costs a fraction of a deep multi-page research run.
- Empty results are free. If the agent finishes without producing a value, you are not charged for it.
Every charge is itemised in the run log next to the answer it paid for, so you can see which prompts are expensive and shorten them.