GPT Image 2.6 API

GPT Image 2.6 is OpenAI's latest image model, released on 4 November 2026 and shipped in ChatGPT as "ChatGPT Images 2.6". The headline is motion: for the first time, an OpenAI image model returns animated GIFs straight from a prompt, turning a mascot, a product shot or a sketch into a short loop. Alongside that come markup-guided edits, SVG vector output and a setting that controls how much the model plans before it draws, plus OpenAI's claim of 50% faster generation.

For developers, OpenAI again splits the release in two, as it did with GPT Image 2.5: Flare, tuned for speed, and Sunburst, tuned for detail and consistent edits. On Unicorn API, GPT Image 2.6 arrives as the Economy tier, a pair of models with a focused option set: gpt-image-2-6-economy-text-to-image for new images and animated GIFs, and gpt-image-2-6-economy-image-to-image for edits and multi-reference compositions.

This article covers where 2.6 comes from, what changed, what the first scores and tests say, what the Economy models on Unicorn can and cannot do, and how to prompt them.

From GPT Image 1 to 2.6

OpenAI's GPT Image line began in March 2025 with GPT Image 1. GPT Image 1.5 followed in December 2025, and GPT Image 2 in April 2026 was the big leap: the first OpenAI image model that reasons before drawing, with far better text rendering. GPT Image 2.5, in September, made it faster and steadier under repeated edits and introduced the Flare and Sunburst split.

Two months later, 2.6 keeps that structure and adds new kinds of output. OpenAI's own summary of the release is short: "Images, upgraded", now with GIFs and reference images, and two API models with 50% faster image generation. The company also says its products now generate four billion images a week, which explains the focus on speed.

GPT Image 2.6 is a closed model. OpenAI has not published weights, a parameter count or an architecture paper. It is available in ChatGPT and Codex on every tier, on web, desktop and mobile, through OpenAI's image generation and edit endpoints, and through API providers such as Unicorn. It is not offered through OpenAI's Batch endpoint, and at launch it has no dated snapshot, only the undated model names.

What changed from GPT Image 2.5

Animated GIFs from a prompt

The most visible addition is native GIF output. Describe a scene with one small movement, a cat flicking its tail, steam rising from a cup, a logo that bounces once, and the model returns a looping animation rather than a still. OpenAI pitches it for characters, mascots, product shots, sketches, reactions and storyboards. GIFs come from the generation endpoint only; you cannot edit an existing GIF or use one as an edit reference yet.

OpenAI's own documentation is not consistent on the frame limit: its guide mentions up to 64 frames while its API reference lists 16. On Unicorn the text-to-image model offers frame_count values of auto, 4, 9, 16, 25, 36 and 64, and frame_duration_ms from 20 to 2,000 milliseconds per frame.

Markup-guided edits

ChatGPT's markup tool, where you circle or scribble over the part of an image to change, comes to the API. A transparent overlay with arrows, outlines or colour blocks tells the model exactly where an edit belongs. OpenAI has since enabled the same feature for the 2.5 models, so it is not exclusive to 2.6, but it arrived with this release.

SVG vector output

Through OpenAI's API, 2.6 can return static SVG files for icons, logos and simple graphics, with editable paths and no embedded raster images. SVG output is not one of the Economy options on Unicorn.

A complexity setting

A new complexity parameter (low, medium or high) controls how much planning the model does before rendering: low for quick edits, high for infographics and crowded scenes. This too is part of OpenAI's full API rather than the Economy models.

Faster generation

OpenAI says image generation is 50% faster than before, and that Flare can return an image "in as little as six seconds" at low quality. Independent testing at CellCog found Flare took roughly half the time of its 2.5 counterpart across quality levels, 18.0 seconds against 45.1 at the highest setting. Sunburst went the other way in the same tests, running 17% to 27% slower than 2.5 Sunburst on identical prompts.

Saved references in ChatGPT

In the ChatGPT app, users can now save reference photos of themselves, pets or characters and reuse them by name without uploading them again. That is an app feature; in the API, you pass references with each request.

What OpenAI has said about safety

Beyond the Flare and Sunburst split, OpenAI has published little about internals. Its safety evaluation reports that the share of unsafe images getting through moved slightly, down for Sunburst and up for Flare compared with 2.5, with neither difference statistically significant. OpenAI also warns that safeguards for animated output "can miss violations involving motion or meaning across frames", a new class of problem that comes with GIFs.

GPT Image 2.6 Economy on Unicorn

The Economy tier keeps the request simple. There is no resolution, quality or complexity setting to tune; you choose the frame, the format and the background, and the model handles the rest. auto aspect ratio lets the model pick a frame that suits the prompt, or the input images in the edit model. In the image-to-image model, input_urls is optional, and with several images you refer to them by order in the prompt ("the first image", "image 3").

First scores and tests

Arena's preliminary leaderboard, published shortly after launch, placed Sunburst second on text-to-image at 1409 ± 13 and Flare fourth at 1384 ± 13, with Sunburst third on single-image editing at 1476 ± 10. Rapidata's preference ranking on 5 November gave Sunburst 1316 ± 13 and Flare 1293 ± 16. Those scales cannot be compared with each other, and the Arena entries are marked preliminary.

Hands-on testing at CellCog was largely positive. Both models spelled every line of text correctly in its prompts, including across GIF frames and edit chains. Sequential edits preserved a character's identity well, with 2.6 holding small details such as a stripe on a hoodie slightly more faithfully than 2.5. GIFs kept characters consistent from frame to frame. The criticisms were specific: GIF loops jumped back to the start rather than closing smoothly, launch-day SVG requests mostly failed with server errors, and the complexity setting made no measurable difference in its tests. Another review pointed out that OpenAI's 50% speed claim comes without a stated baseline or method.

Where GPT Image 2.6 is strongest

Animated stickers and reactions. Short loops of a character waving, a mascot blinking or a product rotating are now one request rather than an animation project.

Text that comes out right. Posters, labels, menus and signs keep the exact wording you put in quotes, and independent tests found no spelling slips.

Edits that keep their subject. Up to 16 input images and steady identity across edit chains make it reliable for changing an outfit, background or label while the face or product stays the same.

Transparent assets. With background set to transparent, icons, stickers and product cut-outs come back as PNGs with a real alpha channel.

Long, specific briefs. Prompts of up to 20,000 characters leave room for a full creative brief: layout, palette, copy and what must not appear.

Limits to know

GIF loops are not seamless. Testers found the animation jumps back to the first frame rather than looping cleanly. Plan for motions that read well as a restart, or keep the movement small.

GIFs only from text. The image-to-image model returns PNG only, so you cannot animate an uploaded photo with the Economy models.

A focused option set. Economy has no resolution, quality or complexity control and no SVG output, and its aspect ratios leave out 3:4, 4:5 and 9:21. If you need those, GPT Image 2.5 Sunburst offers resolution choices up to 4K on Unicorn.

Not first on Arena. At launch Sunburst ranks second for text-to-image and third for editing, so 2.6 is an upgrade in features and speed more than in raw image quality.

Animated content is harder to police. OpenAI's own caveat about safeguards missing problems that only appear across frames is worth taking seriously for user-generated content.

Prompting GPT Image 2.6

Write prompts as a brief: what the image is for, the subject, the setting, materials and light, and any text in quotes. For edits, say what changes and list what must stay. For GIFs, describe one clear, small motion and how it starts and ends.

A still with exact text:

A vertical bakery window poster. Headline "FRESH BREAD DAILY" in warm hand-lettered type at the top, a rustic loaf and a basket of rolls on a wooden board in the middle, and at the bottom the line "Open from 7 am". Cream, rust and dark brown palette. No other text.

An animated loop, with output_format set to gif:

A small round robot mascot sits on a desk and waves its right hand twice, then returns to rest. Flat illustration, soft pastel colours, plain light background, the robot stays centred the whole time.

An edit with references:

Image 1 is the model, image 2 is the jacket. Dress the person in image 1 in the jacket from image 2. Keep the face, hair, pose, background and lighting exactly as in image 1.

A transparent asset, with background set to transparent:

A glossy sticker of a cartoon lemon wearing sunglasses, thick white outline, no background.

Make one change per edit turn and feed the result back as the next input; repeating the details that must stay protects them across a chain.

GPT Image 2.6 against the rest of the catalog

GPT Image 2.5 Flare and Sunburst remain the choice when you need a resolution setting, 4K output or the widest choice of aspect ratios, and they share the same strengths in text and editing. GPT Image 2.6 Economy is the pick when you want animated GIFs or a simpler request.

Nano Banana 2.1 from Google is the main rival for edits and text-heavy layouts and is worth a side-by-side test on posters and product edits. FLUX.3 [klein] is the open-weight alternative from Black Forest Labs, and Seedream 5.0 Flash from ByteDance is another fast option that also offers layer decomposition.

None of those image models return animated GIFs. For a short moving loop from a prompt, GPT Image 2.6 does it in one request; for longer motion, camera moves or sound, a video model is the better tool.

Getting started

Open the GPT Image 2.6 Economy playground to try a prompt, or the image-to-image playground to edit with your own pictures. For code, create a key on the API keys page and send a job to /v1/jobs with the model id, a prompt, and optionally aspect_ratio, background and output_format; add input_urls for edits, or frame_count and frame_duration_ms for GIFs. The docs explain how jobs work, and the API tab on this page has ready-made examples.

Verdict

GPT Image 2.6 is the release where OpenAI's image model learns to move. Native GIFs, markup-guided edits and SVG output widen what a single request can produce, and Flare's speed gain is real in independent tests. It is not a leap in still-image quality: Sunburst sits just below the top of Arena, loops do not close cleanly, and launch-day SVG was unreliable. On Unicorn, the Economy models strip the request down to what matters, the frame, the format and the background, and make animated loops one prompt away.