GPT Image 2.6 團隊The short answer: they are unrelated models from two different companies, and the shared number is a coincidence. GPT Image 2.6 is OpenAI's, released November 4, 2026, and it ships as two IDs: Flare for speed and Sunburst for detail. MAI-Image-2.6 is Microsoft's, a diffusion model whose model card is dated September 4, 2026 and that is built around multi-reference editing and web context. Pick OpenAI's for animated GIF and SVG output on the generations endpoint, plus a complexity dial, and Microsoft's when several reference images must survive one composition.
Two companies, one version number
The names collide, but the two models come from different companies and appear in different documentation.
OpenAI published GPT Image 2.6 in its API changelog on November 4, 2026. Microsoft's model card names Microsoft Corporation as the developer and carries the date September 4, 2026. It covers both MAI-Image-2.6 and MAI-Image-2.6-Flash.
Both are real. Neither is a rebrand of the other. The model IDs are spelled differently and belong to different vendors.
At a glance
One table shows how little the two models overlap.
| GPT Image 2.6 | MAI-Image-2.6 | |
|---|---|---|
| Company | OpenAI | Microsoft Corporation |
| Release date | November 4, 2026 | September 4, 2026 (model card date) |
| Model IDs | gpt-image-2.6-flare, gpt-image-2.6-sunburst |
MAI-Image-2.6, MAI-Image-2.6-Flash |
| How it works | Not described on the model pages | Diffusion with a flow-matching loss |
| Parameters | Not listed | 20B non-embedding |
| Context length | Not listed | 32K tokens |
| Reference images | Not listed on the model pages | Up to five |
| Output size | Presets, or custom in multiples of 16 (max 3:1) | Total pixel cap of 2,359,296 |
| Animated GIF | Generations endpoint only | Not listed on the model card |
| SVG | Generations endpoint only | Not listed on the model card |
GPT Image 2.6 is two models, not one
There is no bare gpt-image-2.6 model ID, so the name alone will not get you a request.
OpenAI splits 2.6 into Flare, the smaller model for workflows where speed matters most, and Sunburst, the base model for detail and precise edits. Both generate, edit, and return transparent backgrounds.
complexity is the new dial. It sets how much the model plans before it renders, and both variants fall back to high when you leave it out.
Drop it to low for one targeted change. quality is a separate axis, so hold one fixed before you compare the other.
MAI-Image-2.6 was built around the reference image
Microsoft's model is built for edits that keep a layout intact.
The Microsoft model page says it creates and edits images from text, images, web context, and visual references. Its multi-reference feature takes up to five visual references alongside your brief.
The model card adds 32K of context, a 20-billion-parameter diffusion model, and edits that hold visual consistency across iterations.
Output is capped at 2,359,296 total pixels, equivalent to 1536×1536.
Output formats are where they split
If you need a moving image or a vector file, only one side documents a path to it.
OpenAI's output_format accepts PNG, JPEG, and WebP. On the generations endpoint it also accepts svg and gif.
The edit endpoint accepts neither, and the docs say so directly.
A GIF defaults to 9 frames at 200ms each, and both the frame count and the per-frame duration are settable. Microsoft's model card does not list a GIF or SVG output path.
Pick by task
Match the task to the model and the choice stops being close.
| Your task | Better fit | Why |
|---|---|---|
| Animated GIF from one prompt | GPT Image 2.6 | GIF output on generations, no listed Microsoft equivalent |
| Editable vector logo | GPT Image 2.6 | SVG output on the generations endpoint |
| One composition from five references | MAI-Image-2.6 | Multi-reference editing up to five images |
| A shot that should follow web context | MAI-Image-2.6 | Web context is a listed input |
| Layout-heavy edit with callouts | GPT Image 2.6 | Markup overlay plus complexity control |
| Fast draft iteration | GPT Image 2.6 Flare | The smaller of the two 2.6 variants |
Markup and complexity are the two settings worth reading up on. Markup takes a transparent PNG overlay at the exact pixel size of your first input image, and it works on GPT Image 2.5 as well as 2.6.
Cost runs on two different meters
The two models do not bill the same way, so cost per image is the wrong unit for a head-to-head.
OpenAI charges $8 per million image input tokens, $30 per million image output tokens, and $5 per million text input tokens. GPT Image 2.5 carries identical rates, as the 2.6 vs 2.5 comparison shows.
GPT Image 2.5 also has a Batch rate of $15 per million image output tokens. GPT Image 2.6 is absent from the Batch table.
Equal token rates are not equal image costs. Two models can spend different token counts on the same prompt, so measure cost per accepted image.
Microsoft ships its model through Azure AI Foundry and the MAI Playground. Its model page carries no public per-token rate.
Running GPT Image 2.6 here
This site runs both 2.6 variants, so you can test Flare against Sunburst before you commit.
Credits scale with resolution. Flare is 20, 30, or 60 credits at 1K, 2K, and 4K; Sunburst is 26, 39, or 78. Animated GIF is a flat 40 or 52.
high quality adds 50%, and low costs the same as medium. Up to 16 reference images are allowed, and failed generations return their credits.
No OpenAI account or API key is needed. Full credit tables sit on the pricing page, the parameter reference lives under GPT Image 2.6 API details, and you can open the generator to try either variant.
FAQ
Is GPT Image 2.6 the same model as MAI-Image-2.6?
No. OpenAI builds GPT Image 2.6 and Microsoft builds MAI-Image-2.6. They share a version number. Different companies, different model cards, different output options.
Which one is newer?
GPT Image 2.6. OpenAI shipped it on November 4, 2026. Microsoft's model card for MAI-Image-2.6 and MAI-Image-2.6-Flash is dated September 4, 2026.
Does MAI-Image-2.6 make GIFs?
Microsoft's model card lists image output only, with no GIF path. GPT Image 2.6 does generate GIFs, but only through the generations endpoint.
Can GPT Image 2.6 edit with a mask?
Yes, and there is a second option. A mask works, and markup lets you send a transparent PNG overlay instead. The overlay must match the pixel size of your first input image.
Do the two models cost the same?
Not in a way you can compare directly. OpenAI publishes token rates, and Microsoft routes its model through Azure AI Foundry without a public per-token table. Measure cost per accepted image on your own prompts.
Sources
- OpenAI API changelog
- GPT-Image-2.6 Flare model page
- GPT-Image-2.6 Sunburst model page
- OpenAI image generation guide
- OpenAI image prompting guide
- OpenAI API pricing
- MAI-Image-2.6 model page
- MAI-Image-2.6 model card (PDF)
Revision log
- November 5, 2026 — Page published.