OpenAI Launches ChatGPT Images 2.5 With Up to 50% Lower Latency, Sketch, Flare and Sunburst
OpenAI has launched ChatGPT Images 2.5, a new generation of its image creation and editing system that the company says is faster, more precise and better at preserving subjects across edits. The update began rolling out on September 8, 2026 across ChatGPT, ChatGPT Work and Codex, while developers received two new API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
The headline performance claim is speed. OpenAI says image generation latency in ChatGPT Images 2.5 is reduced by up to 50% compared with ChatGPT Images 2.0. For developers, OpenAI positions Flare as the default choice for most applications and says it delivers higher-quality output than GPT-Image-2 at 50% lower latency. Sunburst is the premium option, designed for workflows where tighter editing control and production-ready visual quality matter more than raw speed.
But speed is only part of the release. Images 2.5 also adds stronger reference-image fidelity, more reliable targeted edits, better multi-turn consistency, improved handling of complex layouts and transparent backgrounds, plus new ChatGPT creation tools including Sketch, Templates, direct comments on images and prompt sharing.
There is also an important pricing detail for developers: the new 2.5 models are not priced at half the token rate of GPT-Image-2. OpenAI's current API pricing lists Flare and Sunburst at double GPT-Image-2's image and text token rates. The company's own documentation cautions that equal or higher token rates do not translate directly into a fixed per-image multiplier because token consumption varies by model, output size and quality. In other words, “50% lower latency” is a speed claim, not a “50% cheaper” claim.
Key takeaways
- ChatGPT Images 2.5 launched September 8, 2026 and is rolling out across ChatGPT tiers on desktop, mobile and web, as well as ChatGPT Work and Codex.
- OpenAI says generation latency is up to 50% lower than Images 2.0.
- Editing is a major focus: the model is designed to preserve people, products and other reference subjects more faithfully while changing only requested elements.
- ChatGPT adds new creative tools including Sketch, reusable Templates, comments placed directly on images and prompt sharing.
- Developers now get two models: GPT-Image-2.5 Flare for speed and broad application use, and GPT-Image-2.5 Sunburst for higher-control premium workflows.
- API token rates are higher than GPT-Image-2: both 2.5 models list image-output tokens at $30 per million versus $15 for GPT-Image-2.
- The API now supports more production controls, including xhigh and max quality levels, custom dimensions, transparent backgrounds, multiple formats and compression controls.
Why OpenAI is pushing image generation harder now
OpenAI says users are already creating more than 3 billion images per week across ChatGPT Images and the GPT-Image API. That scale helps explain why the company is treating image generation less like a novelty feature and more like a core creative surface inside ChatGPT.
The new release is built around a practical problem that has persisted across generative image systems: producing one attractive image is easier than editing the same image repeatedly without losing the subject, changing unrelated details or drifting away from the original composition.
OpenAI's launch material emphasizes that Images 2.5 is better at preserving the identity and visual characteristics of a reference subject. That applies to portraits, but it is equally important for product photography, fashion, branded assets and any workflow where a user wants to change a background, color, object or line of copy without rebuilding the whole image from scratch.
Reference fidelity and targeted editing are the biggest practical upgrades
For most users, the most consequential improvement may not be prettier first-generation images. It may be the ability to edit an existing image while keeping everything else stable.
OpenAI says Images 2.5 can make more precise edits to selected elements and better preserve the untouched portions of an image. That is particularly useful for tasks such as changing a product's background, replacing text on a poster, adjusting clothing, modifying a single object in a scene or iterating on a campaign visual without re-generating the whole composition.
The company also highlights improved multi-turn editing consistency. In a normal creative workflow, a user rarely stops after one instruction. They might start with a product shot, ask to change the lighting, then swap the background, then update the packaging copy, then request a different crop. Each additional edit creates another opportunity for visual drift. Images 2.5 is designed to reduce that drift and maintain more of the original scene across successive turns.
This is where the release starts to feel less like a traditional prompt-to-image generator and more like an interactive image editor controlled through conversation.
Sketch turns rough drawings into image instructions
One of the most visible additions in ChatGPT is Sketch. Users can draw directly inside ChatGPT and use the drawing as a visual reference for generation or editing. OpenAI says users can invoke it by typing @Sketch.
The feature changes the input model from “describe everything in words” to “show roughly what you want.” That can matter when spatial relationships are difficult to explain — for example, where a logo should sit, how a room should be laid out, the angle of an object or the silhouette of a composition.
For creative work, rough visual direction can be dramatically faster than writing a long prompt that tries to encode position, scale and shape in prose.
Templates, image comments and prompt sharing make ChatGPT more collaborative
OpenAI is also adding Templates for common formats such as posters, merchandise, flyers and product photos. Templates give users a more structured starting point rather than requiring every image workflow to begin from a blank prompt.
Another addition is the ability to place comments directly on an image. Instead of describing an edit with something like “change the object in the upper-right corner,” a user can attach feedback to the relevant part of the image. That should make targeted editing more natural, especially when several visual elements are close together.
Users can also share generated images together with the prompts that produced them. That turns successful generations into reusable recipes: another person can inspect the prompt, adapt it and build on the idea rather than reverse-engineering the result from the image alone.
GPT-Image-2.5 Flare vs. Sunburst
For API developers, OpenAI is splitting the new generation into two models rather than offering only one universal image endpoint.
| Model | Best fit | OpenAI positioning |
|---|---|---|
| GPT-Image-2.5 Flare | Most apps, high-volume generation, creator tools, product experiences, visual search, rapid prototyping | Default choice for most applications; higher quality than GPT-Image-2 at 50% lower latency |
| GPT-Image-2.5 Sunburst | Premium visual workflows, campaign creative, polished product imagery, tighter edit control | Higher-precision option with longer generation time |
The important distinction is that Flare and Sunburst are developer-facing API model choices. In the consumer product, OpenAI presents the experience as ChatGPT Images 2.5 rather than requiring everyday users to choose between those model names.
Flare appears designed for applications where responsiveness matters: social-content tools, product configurators, visual search, fast prototyping and services that may need to generate many images. Sunburst is aimed at cases where the user will tolerate more generation time in exchange for tighter control and a more polished final asset.
API pricing: the new models have higher token rates
The latency claim can be easy to misread as a cost claim, so the pricing deserves its own section.
| Model | Image input / 1M | Cached image input / 1M | Image output / 1M | Text input / 1M | Cached text / 1M |
|---|---|---|---|---|---|
| GPT-Image-2.5 Flare | $8.00 | $2.00 | $30.00 | $5.00 | $1.25 |
| GPT-Image-2.5 Sunburst | $8.00 | $2.00 | $30.00 | $5.00 | $1.25 |
| GPT-Image-2 | $4.00 | $1.00 | $15.00 | $2.50 | $0.625 |
At the published token-rate level, Flare and Sunburst are exactly twice GPT-Image-2 across those listed categories. That does not mean every finished image will necessarily cost exactly twice as much. OpenAI's image-generation guide explicitly warns that per-image cost depends on how many tokens a model consumes for a particular size, quality level and request.
Developers should therefore compare actual response usage rather than multiplying old per-image costs by two. The useful economic question is not simply “what is the token price?” but “how many output tokens does this model use to produce the quality and latency I need?”
More control over size, quality, transparency and format
Images 2.5 also expands the controls available to developers. OpenAI's image-generation documentation lists the usual 1024×1024, 1536×1024 and 1024×1536 sizes, while the 2.5 models can also accept custom dimensions in a WIDTHxHEIGHT format within documented limits.
The new models add xhigh and max quality options on top of lower quality settings. The API also supports output-format and compression choices, plus opaque, transparent or automatic backgrounds. Transparent output requires a format that supports an alpha channel, such as PNG or WebP.
Those controls make the models more useful for real production pipelines where a generated image may need to fit a specific card, banner, product page, ad unit or compositing workflow rather than just look good in a standalone preview.
Complex layouts, transparent backgrounds and style adherence improve
OpenAI says Images 2.5 is better at following complicated visual instructions and producing layouts that combine multiple elements. The company specifically calls out transparent-background workflows and stronger adherence to requested styles.
That matters because image generation is increasingly being used for more than cinematic concept art. Posters, packaging mockups, product cards, infographics, social graphics and merchandise all depend on composition and layout control. The more elements a user introduces, the more opportunities a model has to ignore an instruction, distort an object or move something unexpectedly.
The release is intended to reduce those failures, although OpenAI's own developer documentation still lists limitations.
What OpenAI says still does not work perfectly
Images 2.5 is not presented as a solved image-generation system. OpenAI's current documentation identifies several remaining constraints.
- Complex prompts can still be slow. Some requests may take up to two minutes.
- Text rendering is improved but not flawless. Exact placement and clarity can still fail, especially in dense designs.
- Consistency can still drift. Recurring characters or brand elements may occasionally change across separate generations.
That last point is especially important for commercial use. Better reference preservation and multi-turn editing make the system more useful for brand and character workflows, but users should not assume one perfect reference automatically guarantees identical appearance forever.
Safety and provenance: C2PA plus invisible watermarking
OpenAI says prompts and generated images continue to pass through safety checks. The company also attaches provenance signals to generated images.
Its launch post highlights C2PA metadata, while the Images 2.5 system card says OpenAI is also using Google DeepMind's SynthID invisible watermarking across ChatGPT, Codex and API image generation. C2PA provides machine-readable provenance metadata; SynthID is intended to remain detectable even when visible metadata is lost or stripped.
OpenAI also published adversarial safety evaluation results for Flare and Sunburst. These tests use deliberately policy-violating prompts, so the resulting rates should not be interpreted as normal-user production rates. In the overall evaluation, OpenAI reported “unsafe generation presented” rates of 1.09% for Sunburst and 1.41% for Flare, compared with 1.64% for ChatGPT Images 2.0.
The figures are OpenAI's own safety evaluation, not an independent third-party audit, but they provide useful context for how the company says the new models compare with its previous generation under adversarial testing.
Availability: who gets ChatGPT Images 2.5?
OpenAI says ChatGPT Images 2.5 is rolling out to all ChatGPT tiers across desktop, mobile and web. The new image experience is also available to ChatGPT Work and Codex users.
For developers, both GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst are available through the API. The two-model structure gives developers a clearer choice between lower-latency generation and higher-control premium output.
What the release means
The most important change in Images 2.5 is not a single visual-quality claim. It is the continued shift from one-shot generation toward a full editing workflow.
Reference-image fidelity, targeted element edits, multi-turn consistency, Sketch, image comments, templates and prompt sharing all push ChatGPT toward a place where users can start with a rough concept and iterate on the same asset without leaving the conversation.
For developers, Flare and Sunburst make that strategy more explicit. Flare is optimized around speed and scale; Sunburst around control and polish. The tradeoff is that the new models carry higher published token rates than GPT-Image-2, so teams building high-volume image products should benchmark real usage and output quality rather than assuming faster generation automatically means lower cost.
OpenAI did not publish a single conventional aggregate “image quality score” in the launch announcement. Instead, the company demonstrates the release through capability examples, workflow improvements and customer testimonials. That makes hands-on testing especially important for teams deciding whether the 2.5 models justify a migration from GPT-Image-2.
The bottom line
ChatGPT Images 2.5 is a substantial workflow upgrade centered on faster generation, stronger subject preservation and more controllable editing. OpenAI says latency is up to 50% lower than Images 2.0 in ChatGPT, while Flare is positioned at 50% lower latency than GPT-Image-2 for API developers.
The new Sketch, Templates, image comments and prompt-sharing tools also make the ChatGPT experience feel more like a collaborative visual editor than a simple prompt box.
For developers, however, the pricing nuance matters: Flare and Sunburst list higher per-token rates than GPT-Image-2, even though they can be significantly faster. Teams should evaluate speed, token usage, edit reliability and output quality together.
If the model performs as OpenAI describes in real production workflows, the biggest gain may be less time spent regenerating images after an edit accidentally changes something that was supposed to stay the same.