OpenAI releases GPT Images 2.6: 50% faster, with SVG, GIF and markup-based editing

Image generation usually breaks down in three places: long waits, an edit to one detail rewriting the entire frame, and raster output that works for neither icons nor print. GPT Images 2.6 tackles all three at once. This is an update to the two-month-old model, version 2.5, and it is more practical than flashy.

What changed

The headline number: the model is now 50% faster. In the API, Flare mode at low quality produces an image in about 6 seconds. For interfaces where users wait for the result directly on screen, that is the difference between “please wait” and “almost instant.”

SVG output is now available. This is useful for icons, logos and anything destined for print: vectors scale without quality loss, and designers no longer have to manually recreate a raster image.

GIF support has also arrived. Ask for a loop, upload a photo or illustration, and receive an animation. You can change its speed, edit individual frames and share the result. There is an important limitation: a completed GIF uploaded by a user cannot be edited. The API also does not accept an animated GIF as input; the format works only as output.

Another useful detail for chats is named references. Save a photo of a dog as Pepper, then invoke @Pepper in any chat without uploading it again. For recurring characters and brand mascots, this saves both time and frustration.

New API controls

The complexity parameter accepts low, medium and high, with high as the default. This does not control image quality; it controls how much planning the model performs before rendering. Low is sufficient for targeted edits, while high is better for dense scenes containing many objects. There is no separate charge for it.

The second addition concerns markup-based edits. You send a transparent PNG of the same dimensions as the source, containing annotations, and only the marked area is changed. Imagine a product-card layout where the label on a box must be replaced without touching anything else. Draw an annotation over the box, and the model edits only that area. One caveat: this markup cannot be combined with a mask in the same request, so you must choose one or the other.

How much it costs

Pricing remains the same as in version 2.5, per 1 million tokens:

• input text: $5, or $1.25 from cache;
• input image: $8, or $2 from cache;
• output image: $30.

The cost of a single image depends on its resolution and quality. OpenAI does not publish a fixed per-image price. Budgets therefore have to be estimated using real requests rather than a catalog rate.

What to do with it

If you are building a product with image generation, the update is worth testing immediately because the price is unchanged while speed has improved. Specific steps:

• run your typical requests with low and high complexity and compare the results: low may be enough for half of your use cases, saving time;
• for edits, try the transparent-PNG workflow instead of regenerating the entire image;
• test SVG for icons and logos: vector quality is better assessed by a designer than from a demo;
• if you need animation, account for GIF being output-only and keep source assets in static formats.

Overall, this is not a revolution but a mature step forward: less waiting, more control over edits, and formats that are genuinely useful in production. For products, the most valuable feature may be the ability to change an image precisely without putting the rest of the frame at risk. For developers, it matters that everything is available through the API and in Codex on all plans, allowing adoption without a separate budget approval process.