What they did
The poster used Codex with its built-in image model as the illustrator for an open-source sticker set meant to cover all 11,064 living bird species. Each call got a 3x3 grid of nine reference photos (1254x1254, 418px cells) and was asked to redraw the same nine birds in the same cells with realistic texture, a flat grey background, the bird filling about half the cell and the right number of toes. The output was sliced, cut out with BiRefNet and given a white outline. A separate reviewer scored each sticker against its reference for bill, eye, legs, feet and plumage. When a bird failed, the reviewer's note was fed into the next batch prompt and the redraw was checked against that note. Codex was also used as a second auditor over stickers another model had already passed. Up to 8 sessions ran in parallel.
How it went
Over about a week the run made 2,760 grid batches and 16,339 bird drawings including redraws. 7,676 of the 11,064 species passed the bar; the rest are flagged in the published data. Re-auditing 270 stickers showed the first reviewer model was too lenient (88 disagreements against 23), so the rubric was rewritten. The poster went through the $200 weekly usage limit three times.
Worth knowing
Every generated image left about 7MB in the Codex session logs, and the folder reached 46GB before the poster noticed. If you try this, watch ~/.codex/sessions and clear it. Batching nine birds per image cut the number of calls by a factor of nine.