Skip to content

GPT Image 2.6 Turns GIFs Into a Sprite Sheet Problem

Images 2.6 adds animated GIFs, named reference images, image markup and a planning dial in the API, plus SVG export. Here is what each addition changes in a production pipeline, and where the model still fails.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
Six near-identical ceramic bird figures in a two-row grid on a dark surface, lit by a single raking beam

OpenAI shipped Images 2.6 on November 4, 2026, in ChatGPT, Codex and the API. The headline feature is animated GIFs. The more interesting part for anyone running a production pipeline is how the model gets there, and what else came along with it: reusable named reference images, image markup in the API, a complexity dial separate from quality, and SVG export.

That last one deserves skepticism up front, and I will get to why.

The GIF feature is a frame grid, not an animation engine

OpenAI's launch post includes two sample prompts from OpenArt that produce looping chibi animations. Read them closely, because they reveal the actual interface:

Create a single animation sprite sheet for a seamless looping GIF. ... Exactly 6 sequential frames arranged in a precise 3-column by 2-row grid, read left-to-right then top-to-bottom. Every panel has identical dimensions. Fixed front three-quarter camera, same character, same desk, monitor, outfit, and cozy miniature office background. Clear incremental facial and tear motion; sixth frame transitions smoothly toward the first. ... No captions, no words, no numbers, no logos, no speech bubbles, no panel labels, no borders, no gutters.

Nobody asked for a video. They asked for a contact sheet of near-identical stills with controlled differences, a fixed camera, and a first and last frame that match. The model's job is to keep everything constant except the one thing that should move, and the GIF gets assembled from that grid afterwards.

That changes what you write. A prompt like "make the owl wave" gives the model nothing to anchor on, and you get six unrelated owls. A prompt that names the grid, the frame count, the reading order, the frozen elements and the loop seam gives it a spec it can satisfy. It is the same lesson that shows up in multi-panel and turnaround work: constraint lists beat adjectives.

The loop seam is worth its own line. A GIF that does not loop cleanly reads as broken, no matter how good the frames look. Frame six has to sit one step away from frame one, not at the end of a narrative. Frame count then decides your timing: six frames is half a second at 12fps, a full second at 6fps. Decide the duration first, then work backwards to frames, because adding frames to something that already loops smoothly is easy and fixing a bad seam by generating more is not.

Once generated, the GIF's playback speed is adjustable, and you can revise it afterwards with a text instruction. According to heise, GIF animations that already exist outside this flow cannot be edited the same way, so treat generation as the moment you commit.

GIF is a delivery format, not a default

GIF survived because it plays everywhere without a video tag, supports transparency and loops forever. It is also 256 colors per frame with crude dithering, and the file size grows fast with frame count and dimensions. For a mascot reaction or a short product loop it is fine. For anything with gradients, soft shadows or photographic content, it will band, and the fix eats more size than you saved.

Before you build a GIF pipeline, decide what the asset is for. Chat and social embeds with a hard requirement on autoplay: GIF. Modern web surfaces: animated WebP or AVIF at a fraction of the size. Motion that needs real color depth or audio: video, and accept the poster frame and player chrome.

The generation side does not care. It will hand you frames either way, and the sprite sheet is also a fine source for an APNG or a short MP4 built from six stills. Keep the grid, choose the container separately.

Named references replace re-uploading, not prompt discipline

The second addition is saved reference images. You can name a photo and reuse it as a template in later chats without uploading it again. For character and product work that is a genuine workflow change: the face or the packaging stops living in a folder and starts living in the conversation.

It also introduces a failure mode that everyone who works with reference libraries has met. A named reference is only as good as its name and its version. Three photos of the same person, taken six months apart, in different lighting, all saved as "Maya" will produce drift you cannot see coming, because the model picks whichever one carries the most signal. The discipline that pays off is boring: name by project and role, keep one canonical reference per character or product, and retire old ones instead of letting them accumulate.

Image markup and the complexity dial

Two API changes matter for editing. Image markup, previously a ChatGPT-only feature, now lets a developer mark specific regions of an existing image and describe the change in text. That is the difference between "change the jacket" and "change the jacket, in this rectangle, and touch nothing else". Regional edits have always been the leaky part of iterative work, since a whole-image regeneration re-rolls details you already approved.

The other one is a new complexity parameter that controls how much planning the model does before drawing, kept separate from the image quality setting. OpenAI's guidance is explicit: use low complexity for most editing tasks, medium or high for scenes with many elements or constraints. That separation is more useful than it sounds. Previously the only dial was quality, which bundled detail, fidelity and scene reasoning into one knob, so asking for a cleaner edit also bought you a slower, more expensive reroll of the composition.

Now you can hold detail steady and turn thinking up, or the reverse. For a retouching workload, that is the setting you will touch most.

SVG export is a conversion, not vector craft

OpenAI describes SVG in the API as something that makes "designs ready to use across websites, apps, and print". heise reports the mechanics: the model generates a reference image first, and a second model converts that bitmap into a vector file, which is why SVG output costs more than a plain image.

That ordering tells you what to expect. Tracing a bitmap is good at flat marks, solid shapes and simple silhouettes. It is bad at clean node structure, consistent stroke weights, and anything that has to survive being scaled and edited by a designer: soft gradients, photographic detail, hair, texture. You will get a file with the right outline and an unreasonable number of anchor points.

If the deliverable is a logo, an icon set or an illustration that a human will edit later, generate with a vector-first method instead and keep the SVG as paths from the start. If the deliverable is a visual that happens to need a vector container, an export is fine. We wrote about that split before in vector-first generation with Ideogram, and the release does not change the argument. A trace is a convenience, not a craft path.

Speed, price and the parts that still break

Images 2.6 is roughly 50% faster than its predecessor, and the two API models carry different jobs. Flare, at low quality, can return an image in as little as six seconds and improves layout quality for posters, product imagery and other commercial visuals. Sunburst targets higher detail and holds characters and product features more reliably across editing rounds, at medium or high quality.

Token prices are unchanged from 2.5: $30 per million generated image tokens, $8 per million image input tokens and $5 per million text input tokens. SVG costs more on top because of the second conversion pass. For volume planning, the useful consequence is that the cheaper move is routing, not discounting: fast layouts and ideation to Flare, the final edit and anything with a face or a product on it to Sunburst.

OpenAI still lists the same weaknesses it listed a version ago: text rendering, complex layouts and character consistency. Watermarking stays as C2PA metadata plus an invisible SynthID mark, which matters if your client asks how a synthetic asset is identified. The scale claim is 4 billion images per week across ChatGPT, Codex and the API.

What to retest before you trust it

If 2.6 is about to land in a pipeline that already has prompts tuned for 2.5, a short audit saves a week of confused debugging:

  • Regenerate one existing hero image and compare text rendering. Small type is the first thing to shift between versions, and it is the first thing a client notices.
  • Run one multi-panel or multi-view prompt from your library and check object permanence across panels, not just style consistency.
  • Test the loop seam on a six-frame grid twice: once with "sixth frame transitions smoothly toward the first" and once without, and compare.
  • Take your longest editing chain and run it with low complexity, then medium. Note where the extra planning actually buys something.
  • Check whether your aggregator exposes the new parameters at all. Complexity and markup live in the API, and resellers usually lag a release by weeks.

None of this makes GIFs a mature format or SVG a solved one. It does mean the gap between "a model that draws" and "a model you can run a repeating job on" got smaller, mostly through settings that let you say precisely what should change and what should not.

Sources

Share this guide
X LinkedIn