OpenAI released GPT Image 2.5 Sunburst in September 2026 as its flagship image generation and editing model, alongside a faster sibling called GPT Image 2.5 Flare. The obvious question for anyone already shipping images with GPT Image 2 is whether the new model is worth switching to. A blind panel test run by Everypixel’s production team between 22 and 29 September 2026 gives a two-part answer: Sunburst clearly beats Qwen Image 2, but on believability it could not be separated from GPT Image 2.
What Sunburst is, and what was actually tested
OpenAI’s image generation guide positions the two new models differently: “Choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation.” That matters for reading the results, because the panel tested Sunburst on generation only. Editing, the use case OpenAI puts first, was not part of the study.
Sunburst was run through the Wavespeed endpoint at 2k resolution with quality set to high, and images came back at 2048 px on the long side. OpenAI’s API also offers xhigh and max quality settings for this model, and those were not tested.
How the “anchor ladder” works
The test uses what Everypixel calls an anchor ladder: Qwen Image 2 and GPT Image 2 serve as fixed rungs whose strengths were measured earlier and then frozen, so every new model is placed against the same reference images that earlier models met. Sunburst produced three images for each of 48 commercial scenes (144 in total), and each image was paired once with the Qwen Image 2 image and once with the GPT Image 2 image for the same scene, giving 288 blind pairs.
Thirteen members of the production team (4 in art direction, 5 in content distribution, 4 in quality assurance) rated the pairs, three raters per pair, 864 judgments in all. For every pair they answered three separate questions, with a tie allowed: which image is more believable as a real photograph, which follows the prompt better, and which looks better.
Head-to-head results
| Opponent | Scale | Sunburst wins | Tie | Other model wins |
|---|---|---|---|---|
| Qwen Image 2 | Believability | 58.8% | 20.4% | 20.8% |
| Qwen Image 2 | Aesthetics | 66.0% | 12.5% | 21.5% |
| Qwen Image 2 | Prompt adherence | 40.1% | 49.0% | 10.9% |
| GPT Image 2 | Believability | 36.0% | 33.4% | 30.6% |
| GPT Image 2 | Aesthetics | 43.1% | 27.3% | 29.6% |
| GPT Image 2 | Prompt adherence | 25.2% | 60.0% | 14.8% |
Counting a tie as half a win, Sunburst took 69.0% of believability votes against Qwen Image 2, with a Bradley–Terry interval of 63.4–74.6%. A Bradley–Terry model turns pairwise votes into a probability that one model beats the other; here the 95% interval was estimated with resampling clustered by scene. Sunburst also led Qwen Image 2 on prompt adherence (64.6%) and aesthetics (72.2%). The widest gap was in layout and typography, where it won 47 of 54 believability votes, tied 7 and lost none.
Against GPT Image 2 the picture changes. Sunburst won 52.7% of 431 believability votes, and the interval runs from 46.9% to 58.2%. Because that range includes 50%, the data is consistent with a dead heat: the study does not establish that either model looks more like a real photograph. On prompt adherence (55.2%, interval 51.2–59.6%) and aesthetics (56.7%, interval 50.7–63.1%), both lower bounds sit just above 50%, so Sunburst’s lead there is directional and small.
Rater agreement is also weak. The study reports Gwet’s AC1, an agreement statistic that, unlike Fleiss’ kappa, does not collapse toward zero when one answer such as a tie dominates. For believability against GPT Image 2 it is 0.08, the lowest in the study, meaning the three raters on a pair often saw it differently. Aesthetics agreement against GPT Image 2 is 0.16. The source’s own advice is to treat the believability result as a tie, not a narrow win.
The art directors were least convinced
The most interesting split is by job role. The four art directors gave Sunburst 47.7% of believability votes against GPT Image 2, while content distribution raters (54.2%) and quality assurance raters (55.6%) leaned toward Sunburst. Against Qwen Image 2 art directors still preferred Sunburst, but by less: 60.5% on believability and 55.6% on aesthetics, compared with 80.8% and 85.3% from quality assurance. The role samples are small, but they suggest the people closest to final sign-off are the least impressed.
By category, Sunburst fell below an even split against GPT Image 2 on believability in street and architecture scenes (40.7%) and in fashion and beauty (44.4%). Each category holds only six scenes, so these are pointers rather than firm findings.
What this means if you are choosing
If your pipeline currently uses Qwen Image 2 for photoreal commercial images, this test makes a clear case for moving to Sunburst. If you already use GPT Image 2, realism alone is not a reason to switch: the panel could not separate them on believability, and GPT Image 2 held its ground in street, architecture, fashion and beauty scenes. Sunburst looks like a strong candidate for printed material, packaging text and layouts, where it scored 61.1% against GPT Image 2 on six scenes. At a median of 51 seconds per image (35 to 101 seconds across 144 generations), it suits final frames better than fast exploration.
The full methodology, per-category breakdown and data are in a 288-pair blind test of GPT Image 2.5 Sunburst.
What this test does not tell you
- How Sunburst edits images. OpenAI positions it first for editing precision, but this study covers generation only.
- How Sunburst compares with everything the other models can produce. Each anchor contributed one fixed, frozen image per scene, so the comparison is against those specific outputs.
- Which model wins any single scene. Raters agree weakly on individual pairs, so conclusions hold for the 48-scene corpus as a whole.
- Whether you will get the same frames. The endpoint did not accept a seed, so the exact images cannot be regenerated from the prompts alone.
Sandra Larson is a writer with the personal blog at ElizabethanAuthor and an academic coach for students. Her main sphere of professional interest is the connection between AI and modern study techniques. Sandra believes that digital tools are a way to a better future in the education system.
