ChatGPT Images 2.5 in 30 Seconds
- OpenAI announced ChatGPT Images 2.5 on 8 September 2026. Sharper detail, more precise editing, faster generation.
- Generation latency is down by up to 50% compared with Images 2.0.
- Sketch is new: draw inside ChatGPT and use the drawing as the reference for the final image. Type
@Sketchto open it.- Also new: templates for popular formats, comments placed on the image itself for focused edits, and the option to share the prompt alongside the image.
- Rolling out to all ChatGPT, ChatGPT Work and Codex users, on every plan, across desktop, mobile and web.
- Two new API models: GPT-Image-2.5 Flare (the fast default) and GPT-Image-2.5 Sunburst (slower, more precise).
OpenAI says more than 3 billion images a week are already created across ChatGPT Images and the GPT-Image models in the API. Whatever else Images 2.5 is, it is an update to one of the most heavily used products the company ships.
There are two separate stories in this release. One is the model: more natural lighting, richer texture, and a much better grip on the people in your reference photos. The other is the set of tools wrapped around it inside ChatGPT: drawing, templates, and editing by commenting on the image. The second half is the part you will feel first.
What Is ChatGPT Images 2.5?
Images 2.5 is the new version of the image generation and editing engine inside ChatGPT. OpenAI calls it their state-of-the-art image model and claims progress on three fronts: sharper detail, more precise editing, faster generation.
The short spec sheet:
| Item | Value |
|---|---|
| Announced | 8 September 2026 |
| Name in ChatGPT | ChatGPT Images 2.5 |
| API models |
GPT-Image-2.5 Flare, GPT-Image-2.5 Sunburst
|
| Speed | Up to 50% lower latency than Images 2.0 |
| New product features | Sketch, templates, image comments, prompt sharing |
| Access | ChatGPT, ChatGPT Work and Codex, all tiers |
| Platforms | Desktop, mobile, web |
| Provenance | C2PA metadata + invisible watermarking |
Speed matters more here than the headline suggests. Image generation is a trial-and-error loop: the first result is rarely the one you keep, and you converge somewhere around attempt three or four. Halving the wait means twice as many attempts in the same sitting.
Fidelity to the Reference Photo 🖼️
The clearest improvement shows up when you work from a real photo you already have. Images 2.5 is better at carrying a familiar subject into a new setting, style or composition while keeping them recognisable. Distinctive features survive the transformation, and lighting and texture land more naturally.
Below is one of OpenAI's own examples. Only the clothing changes on a printed childhood photo held up to the camera: a red sweater becomes a white tuxedo with a bow tie. The hand holding the print, the shelves behind it, the curl of the paper and the studio backdrop all stay put.
Original
Images 2.5
Source: OpenAI
The same property is what makes reference-led API workflows dependable. If you generate variations from a product shot, those variations have to stay anchored to the source, and the model's fidelity is what decides whether the pipeline is usable.
Precision Editing: Only What You Asked For
The classic failure mode of image models is collateral damage. You ask for the lamp in the corner to go, the model redraws the whole scene, and everything else shifts a little too. OpenAI says Images 2.5 is better at editing only the region you named, holding the rest steady even with complex subjects and busy backgrounds.
Original
Images 2.5
The result of a "make the bed" instruction. Room, lamps and rug are preserved. Source: OpenAI
Commercially this is the whole ballgame. Updating one product, one background or one line of copy in a campaign asset while leaving the subject, composition and brand treatment untouched is a requirement, not a nice-to-have.
Consistency Across Multiple Turns
The second classic failure mode is drift. By the fifth edit the image has quietly degraded, the instruction from step one has been forgotten, and quality is worse than where you started. OpenAI's claim is that earlier changes are now more likely to survive, and that each new edit builds on the last without eroding quality.
The examples in this section of the announcement are not stills but short videos, each stitched together from dozens of consecutive edits: a rotating cube, a travel infographic built up piece by piece, birthday candles added one at a time.
Sketch: Draw the Idea 🎨
Sometimes the fastest way to explain a layout is to draw it. Sketch lets you draw directly in ChatGPT and use that drawing as the skeleton of the final image.
Type @Sketch in a conversation and a drawing surface opens. Rough out the layout of a room, the silhouette of an outfit, or whatever composition you have in mind, then describe the style and the details you want on top of it. The model turns the rough art into a finished image.
The value is in communicating things that are awkward to write down. "A tall window on the left, a low bookshelf on the right, a sofa in the middle" is a sentence a model can misread in ten ways; the same arrangement takes three lines to draw. No drawing skill required, and the point is not the drawing itself but how close the output lands to the picture in your head.
Templates and Editing by Comment
Staring at an empty prompt box is a real problem, and OpenAI's answer is templates. Pick a format such as "Poster" or "Merch", then fill in the information you need to convey, the design elements and the style. Popular formats like flyers and product photos are covered out of the box.
One of the formats templates are aimed at. Text legibility is noticeably better in this release. Source: OpenAI
The second addition changes the editing loop itself: you can now place comments directly on the image. Click the element you want changed, write "remove this" or "make this blue", and the model applies those notes when you hit send. No more describing "the red vase in the top right" in words.
Third is sharing. When you share an image you can now include the prompt that produced it, so someone else can run the same idea with their own photos and details.
Style and Complex Layouts
OpenAI says the model is better at parsing complex visual instructions and turning them into coherent output. Images that carry real-world information are more accurate, and complex layouts, including transparent backgrounds, are handled better.
Style consistency holding across a dense texture. Source: OpenAI
Transparent backgrounds sound like a footnote and are not. If you produce logos, icons or cut-out product shots that go straight into a layout, not having to key out a background afterwards is real time saved. Lucky Liao of Manus says their evaluations put Flare at two to four times the speed of GPT-Image-2, and calls the improved transparent-background generation a good fit for brand assets, presentations and websites.
Adobe's Matt Chotin confirms the new GPT-Image-2.5 models are available inside Firefly. Higgsfield AI's Axultan Alimkulov puts the emphasis somewhere else: what impressed them most was how well the model understands what not to change.
The API: Flare and Sunburst
Two new models, positioned for different jobs.
GPT-Image-2.5 Flare is the default for most applications. It carries the full set of quality, editing and speed improvements, and OpenAI says it produces higher-quality images than GPT-Image-2 at 50% lower latency. Creator and social content, product experiences, visual search, rapid prototyping and high-volume generation.
GPT-Image-2.5 Sunburst targets premium visual work that needs tighter control across edits. Generation takes longer, precision is higher. Production-ready campaign creative and polished product imagery.
OpenAI's API pricing page lists the same per-token tariff for both:
| Item | Flare | Sunburst |
|---|---|---|
| Image input ($/1M tokens) | 8.00 | 8.00 |
| Cached image input ($/1M tokens) | 2.00 | 2.00 |
| Image output ($/1M tokens) | 30.00 | 30.00 |
| Text input ($/1M tokens) | 5.00 | 5.00 |
| Cached text input ($/1M tokens) | 1.25 | 1.25 |
Source: OpenAI API pricing page, 8 September 2026.
Because the unit rate is identical, the decision is about tokens spent rather than price per token. Sunburst is built for longer generations and tighter control, so in practice its cost per finished image will sit above Flare's. Make Flare the default for anything high volume, and reach for Sunburst where the output ships as-is.
Safety and Provenance
OpenAI says Images 2.5 builds on the existing safeguards, with checks running on both prompts and generated images. C2PA metadata and invisible watermarking continue, so images made with OpenAI tools remain technically identifiable. The evaluations are covered in the system card.
Worth knowing in practice: C2PA metadata is stripped by most tools that re-encode an image or take a screenshot of it, while the invisible watermark survives saving and cropping far better. If you publish generated images, assume they are traceable.
Pricing and Availability
In ChatGPT, Images 2.5 started rolling out on the day of the announcement to ChatGPT, ChatGPT Work and Codex users, on all tiers, across desktop, mobile and web. It reaches the free plan too. What varies by plan is not access to the model but your image generation quota.
Staged rollouts being what they are, it may not appear in your account immediately. Updating the app and waiting a few hours usually settles it.
Flare and Sunburst are available in the API now.
The Takeaway
Images 2.5 is not a redesign. It is the same engine, faster, more faithful, and easier to steer. Of those three, control is the one that will show up in daily use: editing without wrecking the reference photo, and still being on-instruction ten turns later, beats sharper texture by a distance.
The new product features point the same way. Sketch is a channel for compositions that are painful to describe. Comments on the image replace describing an edit with pointing at it. Templates deal with the blank page.
If you want the text-side counterpart, our writeup of GPT-6 Astra covers OpenAI's most recent flagship model.
Which of the three new tools would you actually use? I suspect Sketch is the one people underestimate, and comment-based editing is the one that quietly saves the most time.
AI Generated Content Notice
This blog is entirely generated by artificial intelligence. While AI helps create content, it may still contain errors or biases. Verify critical details before relying on them.
Originally published on Proje Defteri, where this post is kept up to date.
Also on the site: more English posts on AI models, Arduino and IoT, and free browser tools for makers and developers - token counter, LLM cost calculator, LCD and OLED bitmap converters.
Your support means a lot! ✨ Comment 💬, like 👍, and follow 🚀 for future posts!
Top comments (0)