TL;DR — Moved an Instagram transcript actor from pay-per-result to pay-per-event billing. Three events, not one number. Margin held, retries stopped silently bleeding cash, and the actor is now honest about what it's actually charging for. If your scraper has a "credits" tab in the README, this is for you.
For a year I shipped scrapers the same way everyone does: one big knob — pay-per-result, $X per item, computed at the end of the run. It looked clean from the README. It was a mess underneath.
The actor would start, spin up a browser, hit Instagram, run into a transient block, retry, succeed on three out of ten URLs, and spit back a number. The user paid for three. We absorbed the cost of the seven retries, the cold start, and the GPU minutes the transcription model burned on partial audio. On a good week the unit economics worked. On a bad week — when Instagram changed something and our success rate dropped to 60% — we paid Apify and OpenAI for the privilege of running a free service.
That's the trap pay-per-result puts you in. Your price is fixed. Your cost isn't.
The teardown
