Larger side and higher JPEG quality usually increase vision tokens and latency. That effect is an ESTIMATE until a run records ACTUAL usage. Completion max_tokens is chosen from class count (adaptive budget), not a usage prediction.

Only wording & classes change. Output JSON is always the same fixed structure.

{
  "detections": [{ "class_name", "confidence", "bbox": {x_min,y_min,x_max,y_max} }]
}

Pull images from Google Drive. Secrets are encrypted and masked in the list.