ChatGPT Images 2.6 can animate a storyboard frame. Just don't call it video

A chibi office worker sits at a desk, watching a stock chart fall. Over six frames his eyes widen, his face crumples, tears stream down his cheeks, and the sixth frame slides back toward the first so the loop never ends. OpenArt made it for OpenAI's launch of Images 2.6 on Wednesday 4 November, and the prompt is printed under the clip. It doesn't ask for a video. It asks for "a single animation sprite sheet": exactly six frames, arranged in a three-column, two-row grid.

That detail explains most of what the update can and can't do for people who make films. Images 2.6 brings animated GIFs to ChatGPT, saved reference images you can call back by name in later chats, and two new models for the API (OpenAI's pay-per-use developer access), GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, which OpenAI says generate images 50% faster than before. OpenAI's own pitch includes a line aimed squarely at filmmakers: creators can "use animated storyboards to explore character movement before video production." That is a pre-visualization pitch, and on the evidence from launch night it is the honest one.

What shipped, and where

OpenAI's announcement says Images 2.6 is "available globally today in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web." The ChatGPT release notes for the same day are more careful about the headline feature: "GIF creation is rolling out in ChatGPT." So the model is everywhere, while the GIF button may not be in your account yet. In ChatGPT you can turn a prompt or an uploaded image into a loop, change the playback speed and edit individual frames.

The other ChatGPT change matters more for day-to-day work. Saved reference images let you store a photo once and reuse it in later chats by name; OpenAI's example saves a dog as Pepper and then asks for a watercolor of @Pepper without uploading the photo again. Reference images themselves aren't new, since the API's edit endpoint already took them. What is new is that a character, a prop or a face can now persist across a ChatGPT project without being re-uploaded every session, which is exactly where continuity used to slip.

The finer controls live in the API. Frame counts, frame timing and SVG export (vector files that scale cleanly for logos and graphics) are developer features, documented for the two new models, and SVG works only on fresh generations, not on edits. OpenAI says it now serves 4 billion images a week across ChatGPT, Codex and the API, up from the 3 billion The Verge reported in September.

A GIF here is a sprite sheet

OpenAI's image guide sets the GIF range at 2 to 64 frames, with a default of 9 and a per-frame duration of 20 to 2,000 milliseconds. It also says the frame count and frame size must fit "the total request pixel budget." Put plainly, every frame is carved out of one image, so asking for more frames gets you smaller ones. A Japanese ChatGPT tester saw it straight away: 16 frames came back at 267 by 358 pixels, and 32 frames at 217 by 217, each loop running 3.2 seconds.

The 64-frame ceiling is the weakest figure in the launch. It is what the documentation says, and we haven't tested it ourselves. The only paid independent API test we found, 80 generations run on launch night by CellCog, an AI design service that published the results as a blog post, got a hard error at 17 frames telling it to use a number from 2 to 16. Until someone reproduces a 64-frame request through the API, the safe working number is 16.

Creator tests on X, mostly from Japanese and Chinese accounts, point the same way. Testers describe one large still cut into frames and call the motion choppy. One ran frame interpolation on top to smooth the result, another reported flickering backgrounds, and a reply under a partner demo noted drift whenever a character travels across the frame. Short single-beat loops hold up best: a turn, a blink, a punch. That is the territory of an animatic, the rough moving storyboard an editor cuts before a shot exists, and it isn't a substitute for a video model.

Faster, with a catch on the bill

OpenAI's 50% speed claim comes without a baseline, a size or a quality setting. In tests it held for Flare: CellCog clocked Flare at roughly twice the speed of GPT Image 2.5 Flare at every quality level, and a product builder who ran 32 benchmarks on their own face measured 2.6 Flare at high quality 34% faster than 2.5 Flare. Sunburst went the other way in CellCog's run, coming out 17 to 27% slower than 2.5 Sunburst. OpenAI positions Sunburst as the detail and edit-consistency model at medium and high quality, so treat it as the slower, more careful option.

The price looks unchanged. Both 2.6 models list at $8 per million image input tokens, $30 per million image output tokens and $5 per million text input tokens, the same rates as GPT Image 2.5. The number of tokens billed per image has moved, though. For one 1024 by 1024 image at low quality, CellCog was billed 824 output tokens on 2.6 Flare against 196 on 2.5 Flare, about $0.025 against $0.006. At medium quality it was 1,449 against 439. From high quality upward the counts were identical. A developer on OpenAI's own forum posted the same multipliers independently, 4.2 times at low and 3.3 times at medium, and called it a cost multiplier rather than a better image algorithm.

Two more limits sit in the model pages. Neither 2.6 model is supported in the Batch API, where GPT Image 2.5 and GPT Image 2 get half-price rows, or in the Responses API image tool. SVG export is billed twice, once for the reference image and once for the text model that writes the vector file. So the cheap draft pass, the high-volume low-quality iteration many workflows run before committing to a final frame, now costs three to four times as much per image on 2.6 as it did on 2.5.

The partners were ready before the post went up

OpenAI named Picsart, fal, Sable and Higgsfield as early-access customers, quoted OpenArt, Figma Weave and Picsart in the announcement, and OpenArt made its showcase GIFs. Within an hour of the launch threads, Higgsfield posted 13 near-identical demos of what it called GIF sprite sheets, each tagged Made on Higgsfield and paired with a seven-day unlimited offer for new users. OpenArt chief executive Coco Mao gave OpenAI the most quotable line: "GIFs are the native language of social media, and GPT-Image-2.6 lets every OpenArt creator speak it fluently from a single prompt." Day-one demos from launch partners show the feature working on the prompts the partners chose. They are marketing, and they aren't independent reaction.

What it means for AI creators

• Saved references are the everyday win. Store your lead character, a key prop or a location plate once and call it by name across a project, then check every board for drift, because consistency is still the model's claim, not a guarantee.
• Use GIFs as motion tests and animatic beats. A two-second loop of a turn or a reaction is a cheap way to check timing and staging before spending video credits on the shot.
• Keep loops short and frame counts low. Every extra frame shrinks the picture, and long actions drift and flicker in the tests so far.
• Anyone drafting at low or medium quality through the API or an aggregator should price their workflow again. Flare is faster, but at those settings each image bills several times the tokens it did on 2.5, and there is no batch discount yet.

Our take

Our take: animated storyboards are a real gift to anyone planning shots, and saved references fix a daily annoyance. But OpenAI kept the rate card and moved the meter, and creators who iterate cheaply are the ones paying for the extra speed.

Timeline

• Date (2026, UTC): Tue 21 Apr. Event: GPT Image 2 released in the API, with half-price Batch support; Source: OpenAI API changelog
• Date (2026, UTC): Tue 8 Sep. Event: GPT Image 2.5 Sunburst and Flare released; Source: OpenAI API changelog
• Date (2026, UTC): Sep. Event: OpenAI reports 3 billion images a week; Source: The Verge, citing OpenAI
• Date (2026, UTC): Wed 4 Nov. Event: Images 2.6 announced; API changelog and ChatGPT release notes entries; GIF creation rolling out; Source: openai.com; OpenAI developer docs; OpenAI Help Center
• Date (2026, UTC): Wed 4 Nov, 18:00. Event: The Verge reports ChatGPT GIF creation; Source: The Verge
• Date (2026, UTC): Wed 4 Nov, 23:35. Event: OpenAI and OpenAI Developers launch threads on X; Source: X API timestamps
• Date (2026, UTC): Thu 5 Nov, 01:09 to 01:39. Event: Higgsfield posts 13 GIF sprite sheet demos as an early-access partner; Source: X API
• Date (2026, UTC): Thu 5 Nov. Event: CellCog publishes its launch-night test; a developer on OpenAI's forum posts matching token multipliers; Source: CellCog blog; OpenAI developer forum

Sources and what is unconfirmed: OpenAI, Introducing Images 2.6, 4 Nov 2026; OpenAI API changelog, image generation guide, image prompting guide, model pages and pricing page, read 5 Nov 2026; ChatGPT release notes, 4 Nov 2026; OpenAI and OpenAI Developers posts on X, 4 Nov 2026; The Verge, Emma Roth, 4 Nov 2026; CellCog blog, 5 Nov 2026 (vendor content marketing; its token figures are matched by an independent developer on OpenAI's forum, its 16-frame cap is not); OpenAI developer forum thread, 5 Nov 2026; creator test posts on X, 4 and 5 Nov 2026. Unconfirmed: whether the API accepts more than 16 GIF frames, the baseline behind the 50% speed claim, Arena leaderboard scores (we couldn't open the leaderboard), and quotas for GIF creation in ChatGPT. We did not run our own tests.

Your Voice

Quoted from public posts on X and credited by handle. If one of these is yours and you want it removed, message @the__generated on X and it comes down.

Got something to say? Tag @the__generated on X