ChatGPT generates 3 billion images every week, and OpenAI has put a new model on this production line: Images 2.5. Beyond being faster and more refined, what this version really changes is the canvas—sketches can serve as instructions, annotations can be pinned directly onto an image, and prompts can be forwarded together with the image. Generating images is no longer a one-on-one conversation with a model; it's collaboration with people. Image quality is the entry ticket; collaboration is the main event.
What does a 3-billion-image-a-week production line need most?
OpenAI released ChatGPT Images 2.5 on September 8, 2026. The update targets the two most-criticized issues: image quality and editing precision. Faces from reference photos now survive into new scenes, and local edits stop dragging unrelated areas along with them.
Users generate more than 3 billion images per week across ChatGPT Images and the GPT-Image API. 2.5 is aimed squarely at that high-volume production line. OpenAI provides exactly one hard metric: image generation latency is reduced by up to 50% compared to 2.0.
The quality story is more subjective. OpenAI points to more natural lighting, finer textures, and reference-photo subjects that stay recognizable in new scenes. No third party has verified these claims yet. OpenAI hasn't provided quantitative benchmarks.
The real focus is precision editing. Change a product or a piece of background text, and the rest of the image often gets caught in the crossfire. 2.5 modifies only the specific area you point to. Subject, composition, and brand treatment stay untouched.
This pain compounds in multi-turn chats: after several rounds of changes, later edits often overwrite earlier modifications, and image quality degrades as well. 2.5 bets on edits that stick, so each instruction builds on the previous round rather than starting over.
The API splits GPT-Image-2.5 into two tiers, Flare and Sunburst, the latter trading longer generation times for more detailed creative tasks.
The canvas just moved from the finish line to the starting line
The bigger change in Images 2.5 is not the images. A generated picture used to close the task. In this update, it reopens it: the same image can be sketched on, annotated, revised and handed back before anyone calls it done.
ChatGPT image generation used to be one-directional. You typed a prompt, the model drew, and a revision meant typing another prompt from scratch. Images 2.5 removes that one-way path. The canvas becomes a workbench where drafts, notes and prompts pile up on top of one another and get reworked as a stack.
For people who struggle with prompts, OpenAI opened a different door: @Sketch. A rough room layout or clothing silhouette works as the instruction. It does not matter if the drawing is ugly, because the shape itself is the guide.
Blank-canvas anxiety has its own fix. Templates such as Poster and Merch preset the composition, so what remains is filling in details through prompts.
Revision is where small requests cause big damage. Ask for another jacket color and the model may change the pose along with it. Annotations work like comments in Figma: you pin a requirement to a specific region of the image, and the model edits only that region, preserving the pose, composition and brand color around it. This is the “only changes what you ask for” line OpenAI keeps repeating. Technically it is called local editing, and in practice it moves the conversation from the chat box onto the canvas.
OpenAI has also made prompt sharing explicit. Send out the prompt you used, and someone else can replicate the result with their own photos and details. Image creators now double as distributors of recipes.
Seen together on one canvas, image generation becomes a stop along the way rather than the destination. What accumulates is not a finished file but a scene built from sketches, annotations and shared prompts.
The first screen of the announcement is about quality, but the real direction-setter is a few screens later
Image quality is just the first screen of this release; the features that truly drive the direction aren't directly related to pixels.
The first few screens of OpenAI's announcement belong to image quality. The real signal is squeezed into the same paragraph a few screens later: Sketch for rough drafts, pinning annotations onto images, and packaging prompts along when sharing. All three actions answer the same question—after an image is generated, how do people hand it off to each other on the image itself.
Sketching sets a starting point for a generation, so you don't have to begin with a blank canvas and guess words; pinning annotations onto an image inserts "review" into the gap between "drafting"—editorial feedback lands on specific pixels rather than floating in chat logs.
Prompt sharing solves a more subtle problem on its own: why should a well-written prompt be yours alone. When sharing an image, check the option to include the prompt, and the recipient can use the same text with their own photo and play again. A private trick is becoming a tradable asset—the logic is similar to when Lightroom presets and VSCO filters were traded, except this time what's being traded is a text recipe.
The combined effect of these three things is that a single generation no longer ends at the moment the image is produced. Annotations let a single iteration land directly on the image; sharing lets iterations flow from one person's hands to the next; Sketch moves the starting point of the next iteration from a blank canvas to a rough draft. The announcement puts image quality on the first screen, but what drives the direction is a few screens later.
Flare for volume, Sunburst for detail: OpenAI splits the image model API into two tiers
OpenAI's solution for developers isn't a one-size-fits-all model, but two lines—one for speed and one for precision—packaged into separate drawers.
The API introduces two names this time. GPT-Image-2.5 Flare is the default option: OpenAI positions it as a daily choice balancing image quality, editing capability, and speed, with latency (the wait time from request initiation to receiving the image) reduced by up to 50% compared to GPT-Image-2, targeting high-volume tasks like social media content, product images, and visual search. GPT-Image-2.5 Sunburst takes a different path: slower generation in exchange for element-by-element fine-grained control, aimed at creative workflows that need to sweat the details. The subtext behind splitting into two tiers is—high-volume tasks shouldn't be slowed down by precision, so run them cheaply; precision tasks shouldn't be burdened by speed anxiety, so let buyers pay for what they need.
The scale of 3 billion images per week explains why OpenAI is splitting: even a 10% latency optimization translates into real money on cloud compute bills. Flare speeds things up to directly handle high-concurrency batch generation scenarios; Sunburst boosts per-image precision, taking on advertising, design, and e-commerce hero images where one image is worth ten. The two lines are independently priced (pricing not given on the truncated official page), so OpenAI can achieve different profit margins at different tiers.
Compared to GPT Image 1.5 in December: at that time, the API only launched one model, emphasizing editing precision and up to 4x speed improvement. Half a year later, it has become two models, and the change reflects OpenAI's view of market segmentation—developers no longer want "a better model," but "a model for each scenario." This is a shift from selling a single tool to selling tool combinations.
Two signals to watch after launch. The first is Sunburst's share of creative API calls: if call volume steadily climbs within three months, it means the high-precision tier has paying customers and the two-tier strategy holds; if most calls still pile onto Flare, Sunburst may end up as a price anchor (a reference used to make Flare look cheap). The second is Flare's latency numbers. OpenAI claims it's halved compared to GPT-Image-2—this is the vendor's figure and needs third-party verification under identical hardware conditions to confirm it holds across all scenarios. Sunburst's generation time hasn't been given specific numbers by the company; until third-party benchmarks come out, exactly how much "slower" it is remains key to judging whether this product is positioned correctly.
Open ChatGPT first, then decide whether to wait for the API
What everyday users can get their hands on today are mainly the new collaboration tools in ChatGPT; API users need to understand the speed difference between Flare and Sunburst.
If you only use ChatGPT and not the API, the things you can try are already fully rolled out on desktop, mobile, and web: all ChatGPT, ChatGPT Work, and Codex users can access Images 2.5, no additional payment or application required.
OpenAI is straightforward about this release—Sketch, commenting on images, and sharing prompts so friends can replicate with their own photos, all live in the ChatGPT canvas.
If you're a developer, you need to understand one key distinction first: the API launched two models simultaneously. GPT-Image-2.5 Flare takes the route of balancing quality, editing, and speed, essentially sharing the same experience as ChatGPT; GPT-Image-2.5 Sunburst is an extra tier for more detailed creative work, with longer generation times. OpenAI hasn't given specific latency numbers for Sunburst, nor has it published the price difference between the two tiers—this can only be verified after the pricing page is updated.
As for the image quality improvements, the only solid comparison figure is: generation latency reduced by up to 50% compared to Images 2.0. More natural lighting, richer textures, and subjects from reference photos being more easily preserved—all of these come from the company's claims, and no third-party independent verification has appeared yet, so just note them as vendor-stated figures.
In the short term, the main observable actions are concentrated on the ChatGPT side: first, try a reference image generation with a photo of yourself, then do several consecutive edits on the same image to see if the early modifications still hold after several rounds; then use Sketch to draw a rough layout as a draft and have the model generate an image based on the sketch; finally, share a prompt that works well for you and let others use their own photos to try it again. On the API side, the more worthwhile first step is to integrate Flare into your existing workflow for a test run, while keeping an eye on the announcement of Sunburst's latency and pricing.
Upload a photo of yourself or a clear subject, have Images 2.5 place it in a new scene, and compare with the original to see if facial features and clothing details are preserved.
Do 3–5 consecutive rounds of local edits on the same generated image, and check whether the changes from the first few rounds still hold by the last round.
Use Sketch to draw a rough room or poster layout as a draft, and verify whether the model can generate based on the sketch structure rather than only the text you input.
Share a useful prompt via the sharing feature, have a friend replicate it with their own photo, and verify the prompt's portability.
If using the API, first run your existing workflow on GPT-Image-2.5 Flare as a baseline, then wait for Sunburst's latency and pricing to be announced before deciding whether to switch to the precision tier.
Source: OpenAI official release (https://openai.com/index/introducing-chatgpt-images-2-5). This article is compiled based on the vendor's own account; all performance, latency, and editing capability data are OpenAI's official statements, with no third-party independent verification available.