Newsletter
Join the Community
Subscribe to our newsletter for the latest news and updates
OpenAI's GPT Image 2.5 image generation and editing model, released September 8, 2026, offers two API tiers: Flare and Sunburst.
Submit your own product to reach creators and founders looking for the next tool to try.
GPT Image 2.5 is OpenAI's image generation and editing model released on September 8, 2026, packaged as ChatGPT Images 2.5 inside ChatGPT and as two API models, Flare and Sunburst.
GPT Image 2.5 turns text prompts and reference photos into still images, and edits existing images without disturbing untouched regions. Inside ChatGPT it is branded ChatGPT Images 2.5 and went out the same day to every tier, plus ChatGPT Work and Codex. In the OpenAI API it arrives as gpt-image-2.5-flare and gpt-image-2.5-sunburst. It is the successor to GPT Image 2, which shipped on April 21, 2026, and the release focuses on fixing slow generation, subject drift across edits, and unwanted frame rewrites instead of chasing a new headline feature.
GPT Image 2.5 splits its improvements between six model-level behaviours that reach the API and three ChatGPT interface features that do not.
Reference-photo fidelity — Subjects stay recognizable across new settings, styles and compositions when working from a reference photo, with more natural lighting and texture transfer.
Scoped editing — A single request such as replacing a jacket or swapping a backdrop leaves other parts of the frame alone; text copy can be rewritten without the layout shifting underneath it.
Edits that survive the thread — Earlier changes carry forward through long conversations, so a ten-turn retouching chain stays viable instead of re-rolling to an earlier state.
Up to 50% lower latency — Generation time drops by about half against Images 2.0, which lets users iterate during the conversation rather than submitting one prompt and waiting; the early customer Manus reported Flare running two to four times faster than GPT Image 2.
Complex layouts and transparency — Dense layouts hold together and transparent-background generation is improved enough to avoid a separate cutout step for stickers, UI art and overlays.
Grounded real-world content — Infographics, charts, maps and other factual labels render more accurately, according to OpenAI's system card.
Sketch (ChatGPT) — Typing @Sketch opens a drawing surface inside ChatGPT, and the drawing becomes the visual reference for generation; this feature is not part of the API.
Templates and shareable prompts (ChatGPT) — Ready-made scaffolds for posters and merch reduce cold starts, and generated images can optionally carry the exact prompt so other users can recreate the result from their own photos.
GPT Image 2.5 is aimed at teams and individuals who need image output at volume, precise editing control, or accurate rendering of real-world text.
Creative and social teams — Draft many concepts at low quality for fast, cheap iteration, then re-run the winning prompt at high quality for final assets.
Developers building image features — Use Flare for high-volume batch jobs, thumbnails, visual search and in-product generation where turnaround matters; OpenAI recommends Flare as the default for most applications.
Product and brand designers — Use Sunburst for production campaign creative, polished product imagery, multi-turn retouching and anything destined for print, since it prioritizes edit precision over generation speed.
Data and marketing teams — Generate infographics, charts and maps where factual text and layout accuracy are part of the deliverable.
GPT Image 2.5 is commonly used for fast social content, precise photo retouching, transparent-background asset creation, and accurate data graphics.
Social media creators: Produce high-quality images for posts at lower latency than Images 2.0, or use Flare to explore multiple directions quickly.
Product photographers and retouchers: Run multi-turn editing chains in Sunburst where each instruction builds on the previous edit, preserving subject identity and untouched areas.
Designers needing transparent assets: Ask for transparent backgrounds and receive overlay-ready stickers, UI art and overlays without an extra cutout step.
Anyone rendering data graphics: Generate infographics, maps and charts where text and layout stay accurate, per the model's system-card claims.
GPT Image 2.5 keeps the same token rates as GPT Image 2, so per-image cost follows from predecessor measurements. A 1024x1024 image is roughly $0.006 at low quality, $0.053 at medium quality and $0.211 at high quality; portrait and landscape sizes run about $0.006 at low, $0.041 at medium and $0.165 at high quality. The new xhigh and max quality tiers have no published per-image estimate. OpenAI says the GPT Image 2 token calculator does not estimate GPT Image 2.5 token use, so users should measure a representative batch before forecasting a monthly bill.
Yes. OpenAI launched it on September 8, 2026 and rolled it out that day to ChatGPT, ChatGPT Work and Codex on desktop, mobile and web. Both API models, Flare and Sunburst, are also live in the OpenAI API.
The IDs are gpt-image-2.5-flare and gpt-image-2.5-sunburst. Flare is the default for most applications; Sunburst is the higher-precision option for editing-heavy creative work.
No. The model pages list the same token rates as GPT Image 2: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. Per-image costs vary with resolution and quality tier.
Sketch is a ChatGPT-only drawing surface. Type @Sketch, draw what you have in mind, and the drawing becomes a visual reference for generation. It is not exposed through the API.
Images generated in ChatGPT for this family carry C2PA content credentials. Provenance metadata may be handled differently between the chat and API surfaces, so verify your own output if the credentials are important to your workflow.
