gpt-image-2 and the gpt-image-2.5 family: an 8,121-token prompt rendered correctly; from about 8,350 tokens the provider returns no 400 but drops the entire prompt and generates an unrelated image (often a realistic portrait or an interior), with usage.input_tokens_details.text_tokens reporting 8. The documented 32,000-character limit is roughly equivalent for English, but Chinese hits the line at about 11,000 characters.
Recommendations:
- Keep Chinese prompts under about 10,000 characters; repeated character sheets and layout rules take the most space, and character appearance can be carried by reference images instead
- If an image has nothing to do with your prompt, check whether that call’s
text_tokensis a single-digit number - Distill long material with a text model before generating, see Long Prompts
← Back to Live Updates · 📚 Monthly archive