ChatGPT Images 2.5: Flare vs Sunburst, New Features, API and What Changed
ChatGPT Images 2.5 is OpenAI’s latest image-generation upgrade, and it is more than a small quality bump. Released on September 8, 2026, the update focuses on three problems that matter in real creative work: generation speed, reliable editing, and keeping a subject consistent when you make repeated changes.
The release also matters to developers. OpenAI introduced two API models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, aimed at different workloads. Flare is positioned as the default for most applications, while Sunburst targets premium work where precise control matters more than speed.
This guide explains what actually changed, how Flare and Sunburst differ, where ChatGPT Images 2.5 fits into creator and developer workflows, and what businesses should test before moving production image workloads to the new models.
The scale behind the update is already substantial: OpenAI says people create more than 3 billion images every week across ChatGPT Images and GPT-Image models in the API. That makes improvements in latency and edit reliability operationally important, not just cosmetic.
ChatGPT Images 2.5 at a glance
| Area | What changed | Why it matters |
|---|---|---|
| Speed | Up to 50% lower generation latency versus Images 2.0 | Faster iteration and higher-volume workflows |
| Reference fidelity | Better preservation of people and subjects | Fewer unwanted changes during edits |
| Editing | More reliable instruction following across multiple turns | Useful for iterative production work |
| ChatGPT tools | Sketch, templates, image comments and prompt sharing | Less dependence on long text prompts |
| API | Flare and Sunburst models | Developers can choose speed or precision |
| Provenance | C2PA metadata and invisible watermarking continue | Helps identify AI-generated media |
Source: OpenAI’s official Images 2.5 announcement.
What is new in ChatGPT Images 2.5?
The most important improvement is not a flashy demo. It is edit stability. Image generators can often make an impressive first picture and then become frustrating when a user asks for one small change. A new instruction may alter the face, clothing, background, camera angle or composition even when those elements were supposed to stay fixed.
OpenAI says Images 2.5 is better at preserving subjects from reference photos and following editing instructions across multiple turns. That is especially useful for product photography, social campaigns, character-based creative work and any workflow where a team wants to refine an existing visual rather than regenerate it from scratch.
The company also says generation latency is up to 50% lower than Images 2.0. The phrase “up to” matters: real latency will depend on the request, model, image complexity and application. Still, cutting waiting time can materially change how an image tool feels. A workflow with ten iterations becomes far more practical when each revision arrives faster.
Sketch changes how users can prompt images
One of the most searchable new features is Sketch. Instead of describing every spatial relationship in words, users can draw directly in ChatGPT and use that rough drawing as a reference for the final image.
This solves a common prompt problem. Language is good at describing style and intent, but it can be awkward for layout. Saying “put the product slightly left of center, with the headline above it and a circular badge in the lower-right corner” is less direct than drawing those elements in roughly the right places.
Sketch therefore makes ChatGPT Images 2.5 useful for people who are not prompt engineers. A marketer can sketch a flyer. A founder can rough out a landing-page hero. A creator can indicate where a person, prop or title should sit. The sketch does not need to look professional; its value is communicating structure.
The update also adds templates for common formats such as flyers and product photos. Users can place comments directly on images to request targeted changes, and prompts can be shared so another person can reuse an idea with different photos or details.
GPT-Image-2.5 Flare vs Sunburst
The API release is where the update becomes particularly interesting for developers. Rather than offering one model for every job, OpenAI split the new generation into two choices.
| Feature | GPT-Image-2.5 Flare | GPT-Image-2.5 Sunburst |
|---|---|---|
| Primary goal | Speed, quality and scale | Precision and tighter editing control |
| Recommended use | Default for most applications | Premium creative workflows |
| Latency | Lower; OpenAI says 50% lower than GPT-Image-2 | Longer generation time |
| Good fits | Social content, product experiences, visual search, prototypes, high-volume generation | Campaign creative, polished product imagery, detailed edits |
| Best decision rule | Start here when throughput matters | Use when edit precision justifies extra time |
Flare is the practical starting point. OpenAI explicitly calls it the default choice for most applications and says it delivers higher-quality images than GPT-Image-2 at 50% lower latency. That combination is important for apps where users expect quick feedback.
Sunburst is not presented as the universally “better” model. It is a more specialized option for workflows that benefit from tighter control across edits. If a creative team is preparing a production-ready campaign asset or polished product image, waiting longer may be a sensible trade if the model changes fewer unintended details.
Where Flare makes the most sense
For high-volume consumer applications, speed is often part of quality. A technically stronger image model can still feel worse if users wait too long for every variation. Flare appears designed around that reality.
Social-media tools are an obvious example. A user may want five thumbnail concepts, three product backgrounds and several caption graphics in one session. Rapid image prototyping is another. Teams rarely need every early concept to be production perfect; they need enough fidelity to decide which direction deserves more work.
Visual search and product experiences are also notable use cases in OpenAI’s announcement. In these systems, image generation may sit inside a larger workflow rather than being the final product. Lower latency can reduce the total time needed for the application to respond.
When Sunburst is worth the slower generation
Sunburst is easier to understand if you think about the cost of a bad edit. For a casual meme, regenerating is cheap. For a campaign asset that has already passed several rounds of review, an unwanted change can create more work than a slower generation would have caused.
Imagine a retailer has approved the model, product position, packaging and background, but wants to change one label or lighting detail. The valuable behavior is not creative reinvention. It is restraint. The system should alter the requested element and preserve everything else as closely as possible.
That is the type of workflow Sunburst targets. Developers should still test both models with their own assets rather than assuming the premium option wins every case. The right benchmark is not simply “which image looks prettier?” It is how often the model completes the intended edit without creating new corrections.
ChatGPT Images 2.5 for marketers and creators
The new interface tools point toward a broader shift in AI image creation. Prompt writing is becoming only one part of the workflow. Sketches, direct comments, reference images and iterative editing let users communicate visually instead of translating every idea into prose.
For a small business, that could mean sketching a promotional layout, generating a first version, commenting on the exact area that needs revision and then sharing the successful prompt with another team member. For creators, improved reference preservation may make recurring characters and visual identities easier to maintain.
However, teams should avoid treating generated output as automatically production-ready. Brand names, small text, product details, hands, reflections and factual visual elements still deserve inspection. A faster model makes iteration cheaper; it does not remove the need for review.
How developers should evaluate the Images 2.5 API
A useful evaluation should measure the workflow, not just isolated sample quality. Start with a fixed set of prompts and reference images that resemble real user requests. Then compare Flare and Sunburst on the same tasks.
| Metric | What to measure |
|---|---|
| First-pass success | How often the first output is usable |
| Edit fidelity | Whether requested changes happen without unrelated changes |
| Subject consistency | How well people, products or characters remain recognizable |
| Latency | Time from request to usable output |
| Retry rate | How many generations are needed per accepted asset |
| Human review time | Minutes spent checking or correcting outputs |
This is more informative than comparing one beautiful image from each model. A model that costs more per request can still be cheaper operationally if it sharply reduces retries and manual correction. Conversely, a precision model may be unnecessary for bulk ideation.
Developers interested in agentic workflows should also look at our coverage of the OpenAI Agents API and GPT-Live-1 API. Together, these releases show OpenAI expanding beyond text generation into APIs for autonomous work, real-time voice and visual creation.
Safety, C2PA and AI image provenance
OpenAI says ChatGPT Images 2.5 continues to apply checks to prompts and images to reduce harmful outputs. It also continues using C2PA metadata and invisible watermarking to help identify content created with its tools.
C2PA is useful because it can attach provenance information to media, but businesses should not treat any single provenance mechanism as foolproof. Metadata can be stripped when files are processed by other software or platforms. Organizations publishing sensitive or commercial AI-generated visuals should maintain their own asset records as well.
Teams should also establish rules for reference images. If employees upload customer photos, unreleased products or licensed creative assets, the organization should know what its policies allow before those files enter an AI workflow.
Is ChatGPT Images 2.5 available to everyone?
According to OpenAI, Images 2.5 is available across all tiers of ChatGPT, ChatGPT Work and Codex on desktop, mobile and web. Developers can access Flare and Sunburst through the API.
Availability does not mean every interface will look identical at every moment. Product rollouts can be staged, and API usage is separate from consumer ChatGPT access. Developers should check OpenAI’s current documentation and pricing before estimating production costs.
What ChatGPT Images 2.5 means for the AI image market
The bigger story is that image-generation competition is moving from spectacular one-shot demos toward workflow reliability. For professional users, the difficult part is often not creating an image. It is making the seventh revision without breaking the six decisions that were already approved.
OpenAI’s emphasis on subject preservation, multi-turn editing, comments and Sketch reflects that shift. The two-model API strategy reinforces it: one model optimizes the common path for speed and scale, while another targets work where control is worth extra generation time.
That does not automatically make ChatGPT Images 2.5 the best option for every organization. Teams should compare it with the image tools already in their stack using their own products, brand guidelines and editing tasks. The most useful test is the percentage of outputs that reach approval with minimal human rework.
Practical takeaway: Flare or Sunburst?
Start with GPT-Image-2.5 Flare if you are building a general-purpose image feature, social-content tool, product experience, visual search system or high-volume generation pipeline. Its stated latency advantage makes it the logical baseline.
Test GPT-Image-2.5 Sunburst when your users repeatedly edit the same asset and small unintended changes are expensive. Campaign production, polished ecommerce imagery and tightly controlled brand creative are stronger candidates.
For ChatGPT users who are not developers, the most meaningful upgrades are simpler: try Sketch when layout is hard to explain, use image comments for local changes, and take advantage of improved multi-turn editing instead of restarting a generation every time.
Frequently asked questions
What is ChatGPT Images 2.5?
It is OpenAI’s September 2026 update to image generation in ChatGPT, with faster generation, improved reference-subject preservation, stronger iterative editing and new creation tools including Sketch and templates.
What is GPT-Image-2.5 Flare?
Flare is the new API model OpenAI recommends for most image applications. It targets quality and speed, and OpenAI says it has 50% lower latency than GPT-Image-2.
What is GPT-Image-2.5 Sunburst?
Sunburst is the precision-focused API option for detailed creative and editing workflows. It takes longer to generate but is designed for tighter control.
Is ChatGPT Images 2.5 free?
OpenAI says Images 2.5 is available across all ChatGPT tiers. Usage limits can vary by plan. API access to Flare and Sunburst is billed separately according to current API pricing.
Does ChatGPT Images 2.5 support image editing?
Yes. Editing is one of the core improvements. OpenAI says the model follows edit instructions more reliably across multiple turns and better preserves subjects in reference images.
Final verdict
ChatGPT Images 2.5 is a workflow-focused release. The headline speed improvement is useful, but the more important change may be reliability: preserving a subject, making targeted edits and letting users communicate through sketches and comments.
For developers, Flare and Sunburst create a clearer choice between throughput and precision. For creators, the update reduces the amount of prompt gymnastics needed to control composition and revisions. The next question is not whether the models can make impressive pictures. It is whether they can reduce the number of generations, corrections and minutes required to turn an idea into an approved asset.
Sources and further reading: OpenAI: Introducing ChatGPT Images 2.5; The Verge: ChatGPT Sketch.
