ChatGPT Images 2.6: What SVG and GIF Output Deliver and Why Vector Logos Still Need Checks
Images 2.6 brings GIFs to ChatGPT and, through the API, SVG vector files as well. The SVGs work as a quick draft, but a finished logo needs checks first: in one of OpenAI's two sample files, the company name depends on system fonts. If your automations still use gpt-image-1.5, gpt-image-1-mini or chatgpt-image-latest, you have to switch by December 1.
What is actually new in ChatGPT Images 2.6?
The main additions are GIF animations in ChatGPT and SVG files through the API, the interface your own workflows call. SVG is a vector format: shapes and colors stay individually editable, and the graphic scales without losing quality. OpenAI released the API models GPT-Image-2.6 Flare and Sunburst on November 4. According to heise, a German tech news site, Images 2.6 is available worldwide in ChatGPT and Codex on all plans, while the GIF feature is rolling out gradually in ChatGPT.
| New feature | Where | What you should know |
|---|---|---|
| GIF | ChatGPT from text, photo or sketch, API for new generation only | According to heise, existing GIF animations can't be edited this way |
| SVG output | API | New generation only, not edits |
| Mark areas in an image and change them by text | API | No extra fee |
Planning depth complexity |
API | The default is high, which can cause unwanted changes |
| Saved reference images | ChatGPT | Name photos and reuse them in later chats |
| Speed | ChatGPT, API | Around 50 percent faster according to OpenAI. At low quality, Flare is said to need six seconds per image in the best case |
Create a logo with AI: are the SVG files good enough?
As a draft, yes. As a final logo file, not without checks. We didn't run our own generations. Instead, we analyzed the source code of the two sample logos in OpenAI's guide. Both come from the same prompt for a fictional bakery.
| Measurement | Flare | Sunburst |
|---|---|---|
| File size | 4,976 bytes | 4,571 bytes |
| Paths | 12 | 24 |
| Named groups | 5 | 7 |
| Fill colors | 2 | 4 |
| Lettering as live text | no, all paths | yes, 2 text elements |
| Font specified | none needed | Georgia and Arial with fallback fonts |
Three things are good: the files are small, they contain no embedded raster images, and the groups have readable names like loaf or ampersand.
The catch is the type. In the Sunburst sample, the company name is live text set in Georgia. If a device doesn't have that font, the browser substitutes another one and the logo looks different. In Flare, the lettering is converted to paths, so it looks the same everywhere, but you can no longer edit it as text. These two files don't tell you which version you will get: two samples are not a rule.
OpenAI adds its own caveat: with recurring brand elements, exact text placement and strict layouts, the model can still struggle. For an existing corporate design, that is the limit that matters most. See also our guide to brand guidelines for AI.
What should you check before using an SVG?
Five checks are enough to start with:
- Search for text: Open the file in an editor and search for
<text. A match means the font depends on the device. Set your brand typeface in Figma or Illustrator and convert it to paths. - Test it small: Scale the logo down to favicon size. Fine details like steam or wheat ears disappear first.
- Recolor it: Switch it to your brand colors once.
- Inspect the curves: The coordinates in both samples are whole numbers on a grid of 1024 units. That is usually fine for icons, but lettering deserves a look at the letter edges.
- Clear the trademark: An AI sketch doesn't replace a trademark search.
What are the GIFs good for?
Short loops in presentations, emails and social posts, less so fast websites. Through the API, you choose 2 to 64 frames at 20 to 2,000 milliseconds each, and the total has to fit the request's pixel budget.
OpenAI's sample GIF weighs about 1.28 megabytes at 512 by 512 pixels, 9 frames and 1.8 seconds. We measured that on the file itself. The guide mentions transparency for PNG and WebP, not for GIF, and the sample has none.
Our take: for a website header, you are better off converting a GIF like this into a short video. For a slide in a sales conversation, it is fine as it is.
What does it cost through the API?
OpenAI doesn't name a flat price per image. Billing runs on tokens, the model's units of computation. According to the price list, Flare and Sunburst cost $30 per million image output tokens, $8 per million image input tokens and $5 per million text input tokens. The same applies to the 2.5 models. For 2.5, OpenAI states explicitly that equal rates don't mean equal cost per image, because token use varies by model and quality. You can estimate the output cost per image with OpenAI's cost calculator.
There are two traps:
- SVG runs on two rates: First a reference image is created at the image rate, then a second model writes the vector code at its own rate. OpenAI doesn't give a total. Measure it on three real motifs using the
usagefield in the response. - No Batch on 2.6: According to the model page, 2.6 doesn't support Batch, the collected processing mode with a waiting period. The 2.5 models are half price there ($15 instead of $30 per million output tokens). For large image volumes without time pressure, 2.5 can remain the better choice.
What does this mean for you and your workflows?
If you only work in ChatGPT, you don't have to do anything. If Make, n8n or your own scripts generate images at OpenAI, check four things this month:
- Shutdown on December 1:
gpt-image-1-mini,gpt-image-1.5andchatgpt-image-latestwill be removed from the API. OpenAI recommends the 2.5 models as the replacement. - 2.5 or 2.6 decides the rebuild: The official replacement, 2.5, also works through
v1/responses, where the new model name in the image tool is all you need. For 2.6, OpenAI so far only namesv1/images/generationsandv1/images/edits. If you work throughv1/responsesand want 2.6, you have to rebuild the call. - Plan for waiting time: According to the guide, complex prompts can take up to 2 minutes. Set your HTTP step's timeout accordingly, then convert the response from Base64 (a file encoded as text) into a file.
- Set
complexity: According to the guide, lower levels can suit small changes like a background or a sign. The default ishigh.
Then there is labeling. According to heise, the images carry C2PA metadata (a proof of origin inside the file) and the invisible SynthID watermark. We found no metadata block in the two SVG samples. That may come from how they were prepared for the guide, so check it on a file of your own. What companies have to label is covered in our article on AI labeling requirements.
Our take: for icons, illustrations and first logo directions, a test is worthwhile because you get an editable file right away. The corporate design itself stays handwork. If it has to be right, get in touch.
Frequently asked questions
Can I convert an uploaded raster logo to SVG with Images 2.6?
Not directly through the API, because the edit endpoint doesn't output SVG. SVG is only created when generating from a prompt. For an exact copy of a logo, tracing it in a vector program remains the safe route.
Flare or Sunburst: which model should I use?
OpenAI describes Flare for fast iteration and Sunburst for detailed visual work and precise editing. The token rates are the same, but consumption can differ. The best approach is to compare both on two or three real motifs.
Do I need to be approved to use the API?
Possibly. According to OpenAI, GPT Image models may require API Organization Verification in the developer console. Check this before your first test.
Can I keep changing a finished Images 2.6 SVG with the model?
No, not through the edit endpoint. The file is normal vector code without scripts, though, so you can keep working on it in Figma or Illustrator. Alternatively, generate a new version with an adjusted prompt.
Sources
Tell us what you're planning. You'll get an honest assessment within 24 hours.
More articles