GPT Image 2.6 API: 50% Faster Image Generation with Animated GIF Output

The GPT Image 2.6 API is OpenAI’s image generation and editing model, released on November 4, 2026, in two versions: Flare for fast iteration and Sunburst for detailed visual work. Both accept text and image inputs and return images. This release adds animated GIF output, SVG export, markup-guided edits, and a complexity setting, and OpenAI reports that it generates images 50% faster than before.

Flare is built for interactive work and can return an image in as little as six seconds at low quality, with improved layouts for posters and product imagery. Sunburst produces higher-quality images and, at medium or high quality, keeps characters, products, and brand details more consistent across edits. Both versions share the same inputs, features, and OpenAI token rates, so the choice comes down to speed versus detail.

According to OpenAI, GPT Image 2.6 generates images 50% faster than before, and Flare improves layout quality for posters and product imagery. The API adds animated GIF and SVG output on the generation endpoint, plus a complexity setting that controls how much the model plans. Markup-guided edits also reach the API with this release, and they work with GPT Image 2.5 as well.

OpenAI says GPT Image 2.6 generates images 50% faster than before. At low quality, Flare can return an image in as little as six seconds, while higher quality tiers, higher complexity, and dense prompts take longer, up to two minutes for complex requests. For the fastest results, pair Flare with low quality, low complexity, and JPEG output.

Per token, no. Both GPT Image 2.6 versions bill image output at $30 per million tokens, the same rate as GPT Image 2.5. Per image, the lower tiers cost more because they use more tokens: OpenAI’s calculator puts a 1024x1024 image at about $0.025 on low and $0.043 on medium, versus about $0.006 and $0.013 for GPT Image 2.5, while high, xhigh, and max cost the same on both. For high-volume drafts that do not need the new features, GPT Image 2.5 remains the cheaper option.

Yes. Set the output format to SVG on the generation endpoint, and the GPT Image 2.6 API returns a static vector file with editable paths, shapes, and colors, suited to logos, icons, and illustrations. Each SVG request returns one image and does not support streaming or the edit endpoint. The model first renders a reference image and then writes the SVG, so the cost covers both steps.

Complexity controls how much the model plans a scene before rendering, separate from the quality setting that controls detail and fidelity. It accepts low, medium, or high and defaults to high when you leave it out. OpenAI recommends low for most editing tasks, such as changing a background or a sign, and medium or high for scenes with dense text, many elements, or tight constraints. Higher complexity can add latency and unwanted changes, so compare results for your task.

Markup is a transparent PNG of circles, sketches, or notes laid over the first input image, sent with a prompt that describes the change. A mask is an alpha-channel image that guides the edit toward a region, and the model may not follow its exact shape. The two cannot be combined in one request. The markup layer must match the first image’s dimensions, and each is limited to 16,777,216 pixels, with a 50 MB cap on the annotation file.

OpenAI notes that GPT Image models can still struggle with precise text placement, with keeping recurring characters or brand elements consistent across separate generations, and with placing elements exactly in layout-sensitive compositions. Complex prompts can take up to two minutes. Check small text and fine details before an asset ships, and use Sunburst with reference images when consistency matters.

Yes. Atlas Cloud serves its models, including GPT Image 2.5, through one API key with pay-as-you-go billing, and GPT Image 2.6 Flare and Sunburst run on the same account. Switching between versions means changing the model name in your request, so you can draft on one model and finish on another without new credentials.