OpenAI introduced Images 2.6 on November 4, 2026, in ChatGPT, Codex, and the API, and the headline promise is speed: the company says image generation is 50 percent faster than before, and its fast API model, GPT-Image-2.6 Flare, can return an image in as little as six seconds at low quality. The release also adds animated GIF output, SVG export, saved reference images, and a new complexity control for how much planning the model does before rendering. People now create 4 billion images per week across ChatGPT, Codex, and the API, OpenAI says, more than half a billion a day, so a change to the image pipeline touches a very large habit. The question that decides whether to switch is not speed alone. The per-token prices are identical to version 2.5, but version 2.6 spends more tokens per image, so the same picture costs more, and the independent arena scores do not yet show the new models ahead of the old ones.
What actually changed
Images 2.6 ships as two API models with the same names as version 2.5, Flare and Sunburst, and the split is the same: Flare is built for fast iteration, including better layouts for posters and product imagery, while Sunburst is built for detailed work and stronger consistency across edits at medium or high quality. The new capabilities are the same for both. GIFs turn a prompt or a picture into a short animation of 2 to 64 frames, with a default of 9 frames at 200 milliseconds each. SVG export makes designs ready to use across websites, apps, and print. Image markup, already in ChatGPT, comes to the API so teams can mark exactly where a change should land. A new complexity parameter, low, medium, or high with high as the default, adjusts how much planning the model does to meet a request, and OpenAI recommends low complexity for most editing tasks paired with the image quality the workflow requires. Saved reference images let users keep pictures of themselves, pets, and characters ready to reuse without uploading them again. Two things that worked on version 2.5 do not work on 2.6: the Batch API and the Responses API image tool are listed as not supported on the 2.6 model pages. Generated images carry C2PA metadata and an invisible SynthID watermark.
The price did not change. The bill did.
The rate card is unchanged. Text input costs 5 dollars per million tokens, image input 8 dollars, and image output 30 dollars, the same for Flare and Sunburst and the same as version 2.5. What changed is how many tokens each image consumes. From OpenAI's own calculator, version 2.6 spends about 4.2 times more output tokens at low quality and 3.3 times more at medium than version 2.5, so a 1024 by 1024 image at low quality goes from about 0.006 dollars to about 0.025 dollars, while at high quality and above nothing changes, at about 0.053 dollars per image. Image APIs bill per token, not per picture, so a model's token appetite is the price, and a model can hold its rate card and still raise your bill. That is the same trap that decides which flash model is cheapest, a pick that changes with the cache and the workload, and it is why the per-image number, not the per-token number, is the one to compare before switching a production workflow.
| Per million tokens | GPT-Image-2.6 Flare | GPT-Image-2.6 Sunburst | GPT-Image-2.5 |
|---|---|---|---|
| Text input | $5.00 | $5.00 | $5.00 |
| Image input | $8.00 | $8.00 | $8.00 |
| Image output | $30.00 | $30.00 | $30.00 |
| Tokens per low-quality image | about 4.2x version 2.5 | about 4.2x version 2.5 | baseline |
| Low 1024x1024 image, about | $0.025 | $0.025 | $0.006 |
| High-quality image, about | $0.053 | $0.053 | $0.053 |
The arena says the new models are not better yet
LMArena's text-to-image board, checked November 4, 2026, lists the new models as preliminary Medium entries. GPT-Image-2.6 Sunburst (Medium) is second at 1409 points with a displayed interval of plus or minus 13, on 2,967 votes. GPT-Image-2.5 Sunburst is first at 1427 plus or minus 5, on 54,913 votes. The gap is 18 points, and the intervals touch at 1422, so with about 3,000 votes against about 55,000, the honest reading is no measurable quality gain yet, not a loss. The Flare pair tells the same story: version 2.6 Flare (Medium) is fourth at 1384 plus or minus 13 on 2,792 votes, behind version 2.5 Flare at 1399 plus or minus 5 on 53,925 votes, a 15 point gap. In the Image Edit Arena, Sunburst (Medium) is third at 1476 and Flare (Medium) fifth at 1462, and Sunburst leads the Cartoon, Anime and Fantasy category. OpenAI holds all top five text-to-image spots. Artificial Analysis has no version 2.6 entries yet on its image boards, so LMArena is the only independent board with new-model scores, and every one of them is preliminary. Preliminary scores move as votes accumulate, and the Intelligence Index v5 methodology audit shows how a board's construction decides what its scores mean, which is why a 3,000-vote preliminary entry deserves a week of watching before anyone calls it a verdict.
Should you switch?
Switch if you need what is genuinely new: GIFs, SVG export, markup-guided edits, saved reference images, or fast low-quality iteration. Stay on version 2.5 if you generate at high-quality stills, where the cost per image is unchanged, the quality votes still favor the older model, and Batch and Responses calls keep working. The 50 percent faster claim deserves the same caution as any vendor speed number: the announcement names no baseline, quality setting, image dimensions, workload, or measurement method. Three checks before moving a workflow: compare your actual quality setting, because low and medium costs rise while high does not; check whether your calls use Batch or Responses, which version 2.6 does not support; and re-check the arena in a week, when the new entries will carry far more than 3,000 votes. The upgrade is real, but it is a features and speed upgrade, not a quality upgrade, and at low quality it is also a price increase wearing a rate card that did not move.
By ByteHerald Staff. OpenAI, November 4, 2026, introduced Images 2.6 in ChatGPT, Codex, and the API, with GPT-Image-2.6 Flare and Sunburst, 50 percent faster generation, GIF output, SVG export, image markup in the API, and a complexity parameter, and said people create 4 billion images per week across its products. LMArena, November 4, 2026, listed GPT-Image-2.6 Sunburst (Medium) second in text-to-image at 1409 points on 2,967 votes and Flare (Medium) fourth at 1384 on 2,792 votes, behind GPT-Image-2.5 Sunburst at 1427 on 54,913 votes, with Sunburst third in image editing at 1476. Eesel, November 5, 2026, reported from OpenAI's own calculator that Images 2.6 spends about 4.2 times more output tokens at low quality and 3.3 times at medium than version 2.5, while per-token prices are unchanged at 5, 8, and 30 dollars per million tokens.
Comments
Post a Comment