OpenAI's new image AI can hand you a logo as real, editable code

Images 2.6 adds SVG files, animated GIFs and circle-to-edit. Jon's agents ran every feature through the API, and the whole tour cost about a dollar.

Listen to this story
3:37
Live demo The lab notebook: every prompt, setting, time and cost, with the SVGs drawn live
Play with the live page

On November 4, 2026, OpenAI released Images 2.6, its newest image model, as two models in its API: Sunburst, for detailed work and careful edits, and Flare, for speed. The image generation guide lists four new tricks for both: files in SVG, animated GIFs, edits guided by circles drawn on a picture, and a "complexity" dial for how much the model plans before it draws. The price list is unchanged from version 2.5: $30 per million image tokens out.

Jon asked his agents to try all of it through the API, the way his own tools would call it. On November 5 they made 24 paid calls. By the API's own usage figures, the whole tour cost $1.11.

SVG got the most attention. An SVG is a picture written as code: shapes and colors a designer can open, edit and scale to any size without blur. The agents asked for the LEJ logo: the terracotta dot, then "LEJ" in dark ink. Their account had to name a second model to write the code (svg.model); without one, the API answered "SVG generation requires 'svg.model'."

The LEJ logo, drawn by Sunburst as a 537-byte SVG

The file is 537 bytes: one circle and three letter shapes, each with a name. The colors are exact. The letters are a generic bold font rather than the site's own, and the gap is wider than the brand's. A busier SVG, a robot newsroom inside a computer, took seven minutes and came back as 246 shapes in 21 named groups. The model added an Apple logo nobody asked for; the agents deleted its two shapes by hand, which is the point of a vector file.

The SVG newsroom: 244 shapes after the logo came out

Then came six styles from the same theme, robots making the news, on Sunburst at high quality: about 50 seconds and 4 cents each. Unasked, the model wrote fitting text into most of them, from an encyclopedia caption to a chalkboard to-do list.

Victorian engraving
Ukiyo-e woodblock print
Claymation diorama
Isometric pixel art
1970s paperback cover
Photoreal tilt-shift

How to use it

Send requests to the Image API with model set to gpt-image-2.6-sunburst or gpt-image-2.6-flare. The API reference lists the settings:

FeatureSettingWhat the agents saw
SVGoutput_format: "svg", plus svg.model28 s to 7 min
GIFoutput_format: "gif", frame_count, frame_duration_ms9 frames in 31 s
Circle-to-edita transparent markup file on /v1/images/edits18 s
Planningcomplexity: low, medium or high (default)little change on one test
Transparencybackground: "transparent"clean cut-out
Previewspartial_images: 0 to 3, with streamingthree on Sunburst, none on Flare
Sizeup to 3840 pixels on a side4K in 72 s, 10 cents

Flare drew the same photoreal town in 9.6 seconds against Sunburst's 51.6. The guide says a GIF can have 64 frames; the API refused 24 and said 16 is the limit.

A 9-frame GIF made in one call
Circle the cat, ask for a dog: the rest of the room stayed put
Two pictures in, one out: the logo on a clay shirt
A 4K seek-and-find poster

The catch: edits can drift. Removing a second unasked Apple logo from the poster also changed the tablecloth beside it.

What nobody knows yet

What an SVG really costs. The docs say the code-writing model is billed too, but its tokens are not in the usage figures. The docs also name a prompt_rewriter setting without saying what it does.

What to watch

A fix to the guide's GIF frame limit, and SVG pricing written out in full.

Sources