On September 8, 2026, OpenAI released its image generation model "ChatGPT Images 2.5." The company says it has improved the model's ability to carry over the features of people and objects from reference images while editing only the specified areas and leaving the rest untouched. Generation wait times have also been cut by up to 50% compared with Images 2.0. But if you read this update as simply an improvement in image quality and speed, you miss half the point. OpenAI has pushed beyond the one-shot workflow—write a prompt, get one image—and into a production workflow where you give instructions directly on the image and refine it through repeated edits.
On the ChatGPT side, new features include "Sketch," which uses hand-drawn lines as a reference; templates for posters, product photos, and other formats; partial editing via comments placed directly on the image; and sharing with an attached prompt. On the API side, two models are now available: GPT-Image-2.5 Flare, built for fast, everyday use, and GPT-Image-2.5 Sunburst, which prioritizes editing precision. At the same time, the API's standard token pricing is double that of the previous GPT-Image-2. Unless you measure generation time, number of revisions, and final cost on the same task, you can't determine which option is actually better for your workflow.
Images 2.5 puts the emphasis on editing rather than generation
According to OpenAI, more than 3 billion images are created every week across ChatGPT Images and the API's GPT-Image models combined. That figure isn't usage of Images 2.5 alone—it reflects the entire model family spanning both ChatGPT and the API. As usage grows, it becomes just as important that a person's face, a product's shape, or a background layout doesn't fall apart with each revision as it is that the first image looks good.
OpenAI states that Images 2.5 makes it easier than before to preserve the features of people and objects that appear in a reference photo, while confining edits to the specified area. The company also says the model retains prior changes across multiple rounds of editing and limits quality degradation. This is an improvement aimed at work like advertising assets or product photography, where it's clear which elements can change and which must stay the same. That said, OpenAI has not published benchmarks measuring how well faces are preserved or how much quality degrades after editing. Nor does the claimed 50% reduction in wait time come with disclosed conditions—prompt type, resolution, quality setting, or whether the figure represents median or upper-bound latency.
The interface changes point in the same direction as the model improvements. Sketch lets users convey spatial relationships that are hard to put into words—furniture layout in a room, the outline of clothing—through simple lines. On-image comments offer a way to point at a target and instruct "remove this" or "change this color." Templates create a guided process for filling in the information needed for a given use case, while prompt-attached sharing lets other users swap in their own photos or conditions to try it themselves. All of these features are designed to reduce the burden of specifying a finished result through text alone.
Images 2.0 strengthened knowledge, instruction-following, and the handling of complex images that include fine text, and it introduced Thinking mode, which uses web search and other tools to reason before generating an image. The 2.5 announcement extends that trajectory, placing less emphasis on nailing a result in one shot and more on the ability to keep refining an image while preserving the original intent. By combining model performance with interface design, OpenAI has turned the trial-and-error process itself into a product feature.
Flare and Sunburst: splitting speed from precision in the API
On the API side, the entry point differs depending on the use case. For one-off generation or editing, developers use the Image API; for products that refine images within a conversation, they build in the image generation tool of the Responses API. The latter incurs, in addition to the cost of generating the image itself, token charges from the main model that invokes the tool. Looking only at the image model's price sheet won't tell you the total cost of a conversational product.
OpenAI describes Flare as the standard choice for fast, high-quality everyday use, and Sunburst as geared toward production work where editing precision comes first. Sunburst trades generation speed for greater precision. Both models support quality settings of low, medium, and high, plus xhigh and max, with auto as the default. The maximum edge length is 3840px, total pixel count ranges from 655,360 to 8,294,400, and aspect ratio can range from 1:3 to 3:1; resolutions above 2560×1440 are treated as experimental.
Comparing standard pricing on the same per-million-token basis, the price increase is unambiguous.
| Billing item | GPT-Image-2.5 Flare/Sunburst | GPT-Image-2 | Ratio |
|---|---|---|---|
| Image input | $8 | $4 | 2x |
| Cached image input | $2 | $1 | 2x |
| Image output | $30 | $15 | 2x |
| Text input | $5 | $2.50 | 2x |
| Cached text input | $1.25 | $0.625 | 2x |
The standard token pricing for GPT-Image-2.5 Flare and Sunburst is double that of GPT-Image-2 across every category—image input, cached image input, image output, text input, and cached text input. That said, this does not mean the cost of a single image is necessarily double. Output token consumption varies by model and quality setting, and the number and size of reference images also affects input costs. In streaming, where partial images are displayed as they render, each partial image adds 100 image-output tokens.
Flare's claim of 50% lower latency and its 2x pricing are not contradictory—one measures processing time, the other measures price per token; they're simply measuring different things. What production teams actually need to track is the total number of tokens consumed to reach a final image, the number of generation attempts, the number of revisions, and the total time required. This includes workflows that draft at low quality and finish at high quality—all of it needs to be totaled up on real projects.
Safety evaluations improved overall, but not uniformly
The Images 2.5 system card evaluates safety measures using a fixed set of adversarial prompts specifically designed to elicit policy-violating images. The share of violating images that were not blocked and were shown to users was 1.09% for Sunburst, 1.41% for Flare, and 1.64% as the Images 2.0 baseline. The overall figures for the 2.5 series are lower.
However, this evaluation cannot be read as a general-use violation rate. The test wasn't based on typical usage distribution—it was specifically constructed to elicit violations such as violent or sexual content. Automated labeling can contain errors, and sample sizes differ by policy category. The evaluation is also limited to the specific combination of model and safety measures in place at the time of testing.
Results are not consistent across individual policy categories. For sexual content, the share of violating images that went unblocked and were shown to users fell from 2.07% under Images 2.0 to 0.52% for Sunburst and 1.04% for Flare. On political content, however, the baseline of 1.06% rose to 1.69% for Sunburst and 2.54% for Flare; for abuse, the baseline of 2.54% rose to 2.86% for Sunburst and 3.27% for Flare. Moreover, the differences in unblocked violation rates shown to users do not reach statistical significance for any policy category. The test used a two-sided exact McNemar test, and no correction was applied for multiple comparisons.
OpenAI inspects prompts, input images, and generated outputs at separate stages, and continues to apply C2PA metadata and invisible watermarking. The more advanced editing makes it easier to depict people and real-world events, the heavier the burden on provenance information and blocking mechanisms becomes. Still, watermarking is not a guarantee against misuse, and an aggregate safety score does not guarantee safety for any specific use case.
What production teams should measure next
ChatGPT Images 2.5 will roll out across desktop, mobile, and web to all ChatGPT, ChatGPT Work, and Codex pricing plans. On the API, users can choose between Flare and Sunburst, but it has not been disclosed how the two are used differently within ChatGPT itself, whether users get to choose between them, or how this interacts with Images 2.0's Thinking mode.
When deciding whether to adopt this, it's best not to take terms like "high quality" or "fast" from the announcement at face value as your evaluation criteria. Instead, prepare the same reference image and editing instructions, and measure how many rounds of editing it takes before elements you need to preserve—faces, logos—start to degrade. For assets involving text or layout, also track how many times you have to send work back due to typos or misalignment. Even OpenAI's own API documentation notes that complex instructions can take up to two minutes to process, and that limitations remain around precise text placement, recurring characters or brand elements, and tight control over composition.
Even if production time is cut in half, total cost won't necessarily fall if per-unit pricing and the number of revisions both increase. Conversely, if a single edit reliably fixes only the intended area, the resulting reduction in labor could outweigh the price increase. What determines the real value of Images 2.5 isn't how good the first image looks—it's how many revisions it takes to reach a finished product, how much that costs, and how well the original intent survives along the way.