The pay-per-result line is the part I'd want to stress-test. We run a crawler that re-pulls a few dozen job sources every six hours to keep a ~6,700-row index current, and the vast majority of what it pulls on any given pass is something it already has. With a per-1K-results meter that math gets uncomfortable: an incremental poll costs roughly what a cold pull costs, so you end up building a dedup layer on top and paying to discover that nothing changed. Is there a since_id or cursor that only meters what's genuinely new since the last call, or does a repeated query bill in full every time?