gpt-image-2.5-sunburst generating a character sheet (“glamorous woman + full-body turnaround + white background”) consistently returned 400 safety_violations=[sexual] through the API, while the ChatGPT web app rendered it. Failed calls took 42 to 51 seconds versus 46 to 57 for successful ones, so the image was rendered first and then rejected by an output-side classifier. That is why the same prompt passes sometimes and fails sometimes, and why moderation: low does nothing here. The web app passes because its chat model rewrites and expands the prompt first (adding clothing and setting); the API sends it verbatim.
A word ablation, changing one block per variant and running variants in parallel, pinpointed the trigger word (“glamorous”) in 20 calls. Leaving the prompt untouched and adding one phrase, “wearing a beige turtleneck sweater and dark trousers”, passed 3 of 3 with the look and layout unchanged.
The full case, the latency-based layer test, the 20-call ablation log and a general checklist are in the new Safety Rejections page: /en/api-capabilities/image-safety-troubleshooting. The gpt-image-2 overview FAQ and error table have been updated to match.
← Back to Live · 📚 Monthly archive