I’m looking at using an OpenAI vision model to count small items in product photos—think screws, pills, or pieces in a parts tray. The images can be busy, and a one- or two-item error matters more than a slightly awkward description.

Has anyone tested small object counting under those conditions? I’m curious whether results stay consistent across repeated runs, and whether asking the model to mark or list each item helps. If you’ve compared its count with a conventional computer-vision approach, I’d like to hear how that went too.