ChatGPT Images 2.5: sharper edits, ~50% faster, Flare + Sunburst API
OpenAI shipped ChatGPT Images 2.5 (Sep 8): sharper detail, more natural lighting and textures, tighter multi-turn edits that keep subjects and brand treatment intact, and up to ~50% lower generation latency versus Images 2.0. Product side adds Sketch (@Sketch), templates (posters, merch, product shots), in-image comments, and shareable prompts. API splits into GPT-Image-2.5 Flare (default quality/speed) and GPT-Image-2.5 Sunburst (higher precision, longer gen). Rolling out to all ChatGPT, Work, and Codex tiers; Flare/Sunburst live in the Images API.
- Fidelity + edit control: Better reference-photo preservation; edits target what you asked and leave the rest alone across complex subjects and multi-turn chats.
- Latency: Company claim of up to 50% faster vs Images 2.0 (Flare marketed the same for API workloads).
- Creative UX: Sketch as a drawn reference; format templates; comment pins; optional prompt share when you pass an image along.
- API split: Flare for volume/creator/product loops; Sunburst for campaign-grade polish. Early customers cited: Higgsfield, Adobe Firefly, Manus, Runway.
- Safety (company): Prompt/image checks, C2PA metadata, invisible watermarking; points to the system card for evals.
What it means
Bandwagon: this is the gold-standard image model story, not a chat-model crown. The product win is edit discipline plus speed (fix one label without regenerating the whole asset). Decision tree: ship volume on Flare, reserve Sunburst for finals. Latency and fidelity numbers are company claims until you bake-off on your brand kit. Ride the workflow control; keep the receipt on “state of the art.”