GPT Image 2.6 API Makes SVG Logos and Animated GIFs
The GPT Image 2.6 API makes SVG logos and animated GIFs. How to request each piece and the math on what it costs.
The GPT Image 2.6 API now lets you request a vector logo (SVG) or an animated GIF straight from OpenAI's image endpoint, and fix a photo exactly where you mark it. OpenAI announced it on November 4, 2026, with two models, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. What it didn't publish is a fixed price per image, so what a piece costs has to be measured on your first test.
What changed with the GPT Image 2.6 API
On November 4, 2026 OpenAI added two models to the API: Flare, built for iterating fast, and Sunburst, built for detailed compositions and precise edits. Both charge the same rates, and both brought four things:
- A complexity control (low, medium or high; high is the default).
- SVG output.
- Animated GIF output.
- Edits guided by marks.
In other words, you no longer get only a flat image. You can get a vector logo, a GIF, or a photo corrected where you drew on it.

How it works: one endpoint, one parameter
Everything goes through the same image endpoint, and what changes is the output format. For vector, the format is SVG. For animation, the format is GIF. The result arrives as base64, so you decode it and save it as a file.
The setup that matters in practice:
- SVG: the format is
svgandnmust be 1, because SVG comes out one at a time. It exists on generations, not on the edit endpoint, and it doesn't support streaming. As documented, the file is static, with no scripts and no embedded raster images. - GIF: the format is
gif,frame_countdefaults to 9 andframe_duration_msdefaults to 200. Nine frames at 200 milliseconds make a 1.8-second loop. GIF doesn't support streaming either, and the frame size and count have to fit the pixel budget or the request is rejected. - Marks: you send the photo as the first image plus a transparent PNG layer of exactly the same width and height where you drew the area to change. It doesn't combine with a mask.
If you get an error, check the output format against the endpoint first.
What a piece costs: the math you can check
OpenAI charges by tokens, not per image: $30 per million output image tokens. That is $0.00003 per token, or three thousandths of a cent, so 1,000 output tokens cost three cents. The guide itself warns that equal rates don't mean equal cost per image, because what moves your bill is the quality and the size you ask for.
OpenAI doesn't publish a tokens-per-image table for 2.6, but it does for earlier models, which works as a reference only:
- A 1,024-pixel square at low quality used 272 tokens on earlier models: 272 times $0.00003 is about $0.008, roughly eight tenths of a cent.
- On GPT-Image-2 at high quality, the guide lists $0.211 for the same image, about 26 times more.
- A hundred pieces come to about $0.82 at the low reference and $21.10 at the high one.
- If you need three tries to get a good one, multiply by three.
These are figures from other models. The number for 2.6 is the one you measure yourself, from the usage each response reports.
Three example cases (hypothetical)
These are examples, not measured results.
- A taco shop's promo banner. Flare, complexity low, a nine-frame GIF at 200 milliseconds: a 1.8-second loop. You can test ten variations in one sitting. A loop under two seconds grabs attention; it doesn't explain anything.
- A coffee brand's vector logo. Sunburst, SVG,
nequals one. As documented in secondary sources, a reference image is created first and the vector is written from it, so the SVG is a starting point to touch up, not exact geometry. - A store fixing a wrong sign in forty photos. The edit endpoint with the photo, a transparent layer with a circle over the sign, and complexity low.
Who this is not for
Skip the API if you generate three images a month: the ChatGPT app already has the Images 2.6 updates. Skip it if you need a final, trademark-grade logo: an SVG that comes out of an image is a draft. And skip it if you need a fixed price to quote a client, because there isn't one.
What breaks, and what isn't verified
- With SVG, a hundred logos are a hundred calls.
- Complexity high is the default, and OpenAI's guide warns it can slip in changes you didn't ask for during an edit.
- 2.6 doesn't support Batch. A batch on 2.5 at $15 per million tokens would be $30 here: double.
- The "50% faster" and "about six seconds per image" claims come from the maker, with no independent comparison, and the six seconds applies to Flare at low quality.
- A third-party site advertising an image for under a cent isn't an OpenAI price, and user reports of four times more tokens at low and medium quality are unverified.
How to apply it in your business?
Measure before you scale. Send ten test pieces with the same prompt on Flare and complexity low, write down the output tokens of each response, multiply the average by thirty and divide by a million. That is the cost of one piece. Multiply by the tries you need and by the pieces per month, and compare it with what you pay today. With the references above, ten pieces cost about eight cents at the low reference and about $2.11 at the high one, and a thousand pieces at the high reference would be about $211. So set your own cap: have the script stop after the first ten. Every change of model, quality or format moves the tokens and your earlier math no longer holds.
For the wider story, read the full breakdown of ChatGPT Images 2.6 and the animated logo.

FAQ
What is the GPT Image 2.6 API?
It is OpenAI's image API with two models, Flare and Sunburst, launched November 4, 2026, with complexity control, SVG output, animated GIF output and mark-guided editing.
Does GPT Image 2.6 have a fixed price per image?
No. OpenAI charges $30 per million output image tokens and doesn't publish a fixed price per image for 2.6, so your first test measures it.
Can the API make an SVG logo?
Yes, on generations with n equal to one. According to the documentation, the result is a static vector file with no scripts or embedded images, and it isn't available on the edit endpoint.
How long is the default GIF?
Nine frames at 200 milliseconds each, which is a 1.8-second loop.
Conclusion
The real change in the GPT Image 2.6 API isn't that OpenAI draws better: the image stopped being a single format and is now vector, animation and targeted correction. What your piece costs is something only your own ten test pieces can tell you.