OpenAI's Docs Now List GPT Image 2.6 Sunburst and Flare With SVG Output. Price It Per Accepted Asset, Not Per Million Tokens.
OpenAI's API model list now includes two image models: gpt-image-2.6-sunburst, described as image generation and editing for detailed visual work, and gpt-image-2.6-flare, described as the one for fast iteration. An aggregator site lists both as released on November 4. If you generate icons, diagrams or marketing images inside a product, the headline feature is SVG output. The headline problem is that the pricing is per token, and you do not buy tokens, you buy assets you keep.
What the docs say
From OpenAI's model and pricing pages, as I read them: both 2.6 models list image output at $30.00 per million tokens, image input at $8.00 per million, and text input at $5.00 per million. One of OpenAI's image guides says both models support complexity control, markup-guided edits, SVG output and animated GIF output through the Image API. Complexity takes low, medium or high, defaults to high, and higher settings can raise latency.
I read the pricing and guides through truncated page views, so treat the exact figures as needing a second look before you budget on them.
What I could not find
There is no per-image price table in what I could read. OpenAI's pricing page points to a cost calculator that I did not find. I also found no OpenAI announcement post and no support for a claim I saw elsewhere that images got 50 percent faster. And the docs disagree with themselves: the guide for the image generation tool still shows the older 2.5 model IDs and lists only png, webp and jpeg output, while the other guide describes SVG for 2.6. Until that is cleaned up, I would not assume SVG works the same in every endpoint.
Why token pricing makes this awkward
An image priced per million output tokens has a cost that depends on size, quality and how complex the picture is, none of which you can estimate from a single sample. Add the fact that you will reject some outputs and re-run them, and your real unit cost is the total spend divided by the assets you actually ship. A cheap-looking model with a 40 percent rejection rate costs more than a pricier one that gets it right.
The test I'd run
- Pick 20 assets you really need: a set of app icons, a few diagram styles, a header image or two.
- Run each through Flare and Sunburst at the complexity levels you would use in production, and log the token usage the API returns for each call.
- Mark each result accepted or rejected honestly, including SVGs that render but are messy to edit.
- Compute total cost divided by accepted assets for each setup.
- Open the SVGs in a browser and in your editor of choice. An SVG that looks right but is a thousand tangled paths is not usable.
Do that before you replace an existing image pipeline, because the answer may be that your current tool is cheaper for icons and this one is better for detailed editing.
Where I could be wrong
Costs may come out lower than my caution implies, and SVG generation might simply be good. OpenAI's own wording about Flare being for fast iteration suggests a cheap-and-quick path exists. The only way to know is to run your own assets.
What I'd actually do
Wait for the cost calculator or the announcement post to appear, run the 20-asset test, and put the cost per accepted asset in a spreadsheet next to your current approach. If SVG output holds up for icons, that is a real saving on design time. If it does not, you have lost an afternoon.
Author
Lukas
@lukcombinator