GPT Image 2.6: what's new, what it costs and how it does on real ads

GPT Image 2.6 is the newest model in OpenAI's GPT Image line, following GPT Image 2.5 in September 2026 and gpt-image-2 before it. It takes text and images in and returns an image, so it can make a picture from a prompt or edit and combine pictures you upload. The API model IDs are gpt-image-2.6-sunburst and gpt-image-2.6-flare, and they work on the Image API's generations and edits endpoints only: unlike 2.5, they are not available through the Responses API or the Batch API.

Like 2.5 before it, 2.6 comes as a pair with the same prices and features. OpenAI calls Flare the smaller model, for workflows where speed matters most, and Sunburst the base model, for more detail or precise edits. Our full comparison is in GPT Image 2.6 Sunburst vs Flare.

If you still use an older model, note the deadline: OpenAI shuts down gpt-image-1.5, gpt-image-1-mini and chatgpt-image-latest on 1 December 2026 and names the 2.6 models as their replacement.

These are the changes OpenAI lists for the 2.6 models. None of them changes the token price.

• Complexity. A new setting that controls how much the model plans before it renders, with three levels: low, medium and high, the default. OpenAI recommends lower levels for focused edits and higher levels for dense text and complex layouts, and warns that higher levels can add time and unwanted changes. It carries no separate fee.

• SVG output. Set output_format to svg on a generation request and the API returns one static vector file. Edits do not support SVG.

• Animated GIF output. Set output_format to gif on a generation or an edit, with 2 to 16 frames (9 by default) shown for 20 to 2,000 milliseconds each (200 by default).

• Markup-guided edits. You draw on a transparent layer the same size as the first image and send it with an edit request, so the model knows exactly where to change things. This works on the 2.5 and 2.6 models.

• Speed. OpenAI says both models generate images 50% faster than before, and that Flare can return an image in as little as six seconds at low quality.

• Provenance. OpenAI's system card says images carry C2PA metadata and a SynthID watermark.

Quality still runs from low through medium and high to xhigh and max, plus auto. Custom sizes need both sides divisible by 16, no side longer than 3,840 pixels and a shape between 1:3 and 3:1.

Both versions cost the same per token. In our tests they cost the same per image too, within a tenth of a cent, because Flare returned about 3% more output tokens. The real difference was time.

OpenAI charges by the token, not by the image: $8 per million image input tokens, $5 per million text input tokens and $30 per million image output tokens. There is no Batch discount for 2.6 yet. The price of one picture depends on how many reference images you send and how large and detailed the output is.

In our tests a 1088 x 1360 ad at high quality used about 1,600 output tokens, roughly $0.048 of the bill. Five reference images added $0.04 to $0.06 of input. A single-image edit came to $0.060. The full breakdown, with a worked example and what 100 ads cost, is in GPT Image 2.6 pricing.

We remade a list-style ad ("DO THIS / NOT THIS") for Virlo, a short-form video analytics tool, using the reference ad and four screenshots of Virlo's dashboard as inputs. The prompt, images, size (1088 x 1360) and quality (high) were identical for each model. We ran a second remake with four product photos of a supplement bottle, to test packaging accuracy.

All three models spelled every line of copy correctly, kept Virlo's real dashboard figures legible on the laptop screen and copied the supplement label word for word, small print included. The differences were in layout and style.

Flare followed the reference ad most closely. On the dashboard ad it kept the laptop in the bottom left with the button beside it, as the reference does, and on the packaging ad it reproduced the reference's outlined lettering. Sunburst 2.6 set bigger, bolder type, and both Sunburst versions stacked the laptop under the lists. None of the three changed the product's colour, a fault we saw on GPT Image 2.5 Flare in September.

If you use GPT Image 2.5 today, the switch costs nothing extra and Flare 2.6 gives you the fastest renders we have measured from OpenAI. OpenAI's "50% faster" held for Flare in our tests, but at high quality with reference images Sunburst 2.6 was slower than Sunburst 2.5 on launch day. If speed matters more than OpenAI's detail and consistency pitch, pick Flare. Our side-by-side of the two generations is in GPT Image 2.6 vs 2.5.

On the public Arena leaderboards of 2 November 2026, OpenAI's GPT Image models held the top five places for text to image, with 2.6 Sunburst second on early votes at medium quality and Google's Nano Banana 2.1 seventh. GPT Image 2.6 vs Nano Banana 2.1 compares the two directly, and the best AI image models for ads covers eleven models for ad work.

The models take custom sizes, so you can ask for Meta's exact ratios. Both sides must be divisible by 16, and 1080 is not, so we request 1088 x 1360 for a 4:5 feed ad, 1088 x 1088 for 1:1 and 1152 x 2048 for 9:16, then scale down to 1080 x 1350, 1080 x 1080 and 1080 x 1920. Put the reference ad first among the input images and say plainly that its brand, product and words must not be copied.

The step-by-step version, with the settings and prompt structure we use, is in how to make Facebook ads with GPT Image 2.6. If you would rather not write prompts at all, Mockingbird does the whole process from a winning ad and your product photos, and renders it on GPT Image 2.6.