Four Billion Images a Week, and Now They Move

OpenAI shipped Images 2.6 on November 4, and the demo everybody shared was a chibi office worker crying at a plunging stock chart. Six frames, one sprite sheet, a seamless loop. Cute, mildly brutal, extremely shareable.
The number buried under the GIFs is more interesting. People now generate about four billion images a week across ChatGPT, Codex and the API. That is over half a billion a day. Eight weeks ago, when Images 2.5 launched, the figure was three billion. Nothing about the underlying model got three times better in eight weeks. What changed is that image generation stopped being an event and became a reflex.
What actually shipped
Three things, and only one of them is about the model.
Speed. Generation is roughly 50% faster. The API splits into two variants: GPT-Image-2.6 Flare, which at low quality can return an image in about six seconds, and GPT-Image-2.6 Sunburst, which is the one you pick when a subject has to survive several rounds of editing without drifting.
Motion. GIFs are now native. You type a prompt or drop a still, and you get a looping animation. Playback speed is adjustable, individual frames can be edited inside ChatGPT, and the result posts directly. OpenAI's suggested uses are mascots, reaction memes, and animated storyboards before anyone books a video shoot.
Editing as an interface. This is the part I keep coming back to. The API now supports image markup, so you can point at a region and say "change this." Images export as SVG. There is a complexity parameter that controls how much planning the model does before it renders, which is a strange thing to expose until you realise it is a latency dial for people who edit in loops rather than one-shot.
Add saved references — you can stash a photo of a person, a pet, or a recurring character and reuse it later without re-uploading — and the shape of the product becomes clear. This is not a generator with extra features. It is a design tool that happens to be driven by a model.

The GIF is the least interesting part, and also the most important
Every outlet led with GIFs, and OpenAI clearly wanted that. Coco Mao, CEO of OpenArt, is quoted saying GIFs are "the native language of social media," which is the kind of line that exists to be put in a press release.
She's not wrong, though. A still image is a post. A loop is a format. Once a model returns a loop from the same prompt box, the output competes with video, not with illustration — and it competes at a fraction of the cost and time. For a social team, the practical question stops being "do we have a video budget" and becomes "does this idea need sound."
The interesting bit is what the loops are made of. OpenAI's own examples are sprite sheets: six frames, three columns by two rows, identical dimensions, fixed camera, frame six easing back toward frame one. That is not how a video model thinks. It is how an animator thinks. The prompt is doing production design, and the model is filling in cells.
Which means the skill that got more valuable this week isn't prompt writing. It's knowing what a loop needs to look like so it doesn't jitter when it restarts.
Four billion a week changes the bottleneck
When you generate forty images for a campaign, the hard part is making them. When you generate four thousand, the hard part is finding the eleven that are usable, and remembering which eleven.
Two of the shipped features only make sense in that world. Saved references exist because consistency across a series is now the constraint, not novelty. Image markup exists because fixing a bad hand is cheaper than regenerating a good picture. Neither is a creative feature. Both are review features.
I've seen this pattern before, in text. The moment output volume crossed a threshold, tooling shifted from "help me write" to "help me find what I wrote." Image tools are on the same curve, roughly two years behind.
The API split tells you who OpenAI thinks is buying
Flare and Sunburst are not a good/better ladder. Flare is tuned for throughput and layout: OpenAI specifically claims better composition for posters, product shots, and other professional visuals. Six seconds at low quality is a number you quote to a product manager who wants thumbnails in a grid.
Sunburst is tuned for consistency at medium and high quality, which is a very different promise. Consistency is what you need when the same character appears in twelve panels, or when a brand kit has to survive an edit pass by someone who isn't a designer.
Put together, the two variants describe two jobs that used to be one: explore cheaply, then commit expensively. Anyone who has run a design sprint will recognise the shape. It's the same reason agencies do fifty thumbnails and one comp.
What I'd actually test
If you have a pipeline, the honest test is not image quality. It's whether the boring parts got shorter.
Take a real brief — a product page hero, say — and time three passes: the first draft, the round of edits after the client says "warmer," and the export. In my experience with previous generations, the middle pass was always the killer, because every edit risked changing the product shape. Sunburst plus markup plus references is aimed exactly at that.
Second test: the loop. Ask for a seamless GIF of something with a moving part, and check frame six. If it snaps back with a visible jump, the model is still thinking in stills and stitching. If it eases, it's actually planning the cycle.
Third test, and the one most teams skip: the archive. Ask yourself where those four thousand images live, who can search them, and whether the ones you rejected are still billable. Volume has a way of turning a creative problem into a records problem.
The uncomfortable part
Watermarks and provenance get harder as output gets cheaper. Four billion images a week is a lot of surface area, and a GIF that loops is more convincing than a still, because motion reads as evidence. The same property that makes loops good for mascots makes them good for fakes.
OpenAI's own framing is careful here: the examples lean on brands, mascots, product shots, storyboards. Not news, not people saying things. That's a sensible line to draw voluntarily, because the alternative is drawing it under pressure later.
So what
Images 2.6 is not a leap in image quality. Read the changelog and you'll find speed, loops, and a pile of editing affordances. The interesting claim is structural: image generation is being folded into the part of the work where people make decisions — mark up, revise, export, reuse.
That's a less exciting story than a crying cartoon office worker. It's also the one that changes how a team spends its week. The four-billion number isn't a brag about scale. It's a warning that the scarce thing is no longer the picture. It's knowing which picture.
Related articles
AMD Got 19 Percent for Being the Other Option
Money did not leave AI. It left the one name with a governance problem and went to the two without one.
The $99 Billion Loop Nvidia Built Around Its Own Customers
Selling shovels is a good business. Lending money to the people buying shovels is a different one.
The 74 Percent Gross Margin Has a Shelf Life
Three forces arrive in the same four quarters, and each one costs about $3 billion per point.
175,000 Signatures Will Not Put Him Back. Here Is What Would.
The longer the campaign takes to win, the weaker its case becomes. That is the trap built into it.