GPT Image 2.5 API in practice: five hidden settings and how each performed
Practical_Low29 · reddit · 2026-09-16
A hands-on test of the GPT Image 2.5 API, documenting five settings the ChatGPT app never exposes:
- Model choice: the API lets you pick gpt-image-2.5-flare or gpt-image-2.5-sunburst; the chat app and Codex built-in tool cannot.
- Five quality steps, not a toggle: low/medium/high/xhigh/max plus auto. On a 1024 product shot, flare ranged from 11.6s (low) to 42s (max); sunburst from 18.7s to 82s. Always set the step manually when comparing outputs.
- Almost any size: width/height multiples of 16, long side ≤3840, ratio ≤3:1, pixel count between 655,360 and 8,294,400. 1024x1536 and 1920x1080 both returned faster than a 1024 square.
- Transparent background is a parameter: background=transparent with png/webp gave 84.5% transparent pixels on a logo; writing "transparent" only in the prompt produces a fake checkerboard painted into RGB pixels — the source of common chat-app complaints. Add "no plate" when the subject usually sits on a plate, or the model draws one.
- Up to 16 reference images with a mask: the edit endpoint supports them. The author's product workflow: flare edit for cutout, flare edit to place into a scene, sunburst edit for retouch, with a "Keep exactly" protection paragraph locking product geometry, labels and packaging text.
More from coding & agent
- Microsoft's MemSteer turns unlabeled user histories into agent memory supervision — MicrosoftResearch · 2026-11-02
- She fired herself from her own Instagram: AI agent runs the account, views up 82.3% — heyshrutimishra · 2026-11-02
- Sourcegraph: AI writes code faster than teams can review it, driving more outages — AI Engineer · 2026-11-02
- Tsinghua & Microsoft's HarnessMM cuts agent token use by up to 87.93% while raising success rate — rohanpaul_ai · 2026-11-02
- GitHub launches Agent HQ: Anthropic, OpenAI, Google coding agents in Copilot — emmanuelvivier · 2026-11-02
- DeepSeek open-sources Super-Agent, a coding agent that reads team chats to minimize human input — emmanuelvivier · 2026-11-02