Six seconds is the number OpenAI is leading with. The company released Images 2.6 on November 4, and in its fastest configuration the model can produce a picture from a prompt in roughly that long when quality is set low, a pace the company says is 50% quicker than the version it replaces.
The model is available worldwide and runs inside ChatGPT, Codex and the company’s API. Beyond speed, OpenAI says the release improves control over detail across multi-turn editing and holds characters, products and brand elements more consistently from one revision to the next. It also adds GIF generation, which turns a still image, a character or a product shot into a looping animation.
The API is split in two. GPT-Image-2.6 Flare is the fast variant, the one built for iteration and volume. Sunburst emphasizes higher quality and consistency across edits, the property that matters when an image has to survive twenty rounds of revision by people who are not artists. Developers also get annotation-based editing, SVG export and a complexity parameter that lets them trade fidelity against cost and latency.
Those features are aimed at a different buyer than the one that made image models famous. The first wave of adoption was measured in astonishment, in pictures that had never existed before. The second wave is measured in approvals, version histories and whether the same character looks like the same character on page four as on page one. That is dull work, and it is where the money is.
Consistency is the hard problem in commercial image generation. A brand asset is not a single image but a family of them: the same product at different angles, the same mascot in different poses, the same label at different sizes. Early models could produce any one of those convincingly and struggle to produce the set, because each generation started from scratch. Multi-turn editing with detail control is an attempt to make the model remember what it decided.
The GIF feature points at the same commercial logic. A looping animation is the standard unit of paid social advertising and a routine deliverable for any brand team. Producing one used to require either a motion designer or a separate tool that took a still image and applied a canned effect. Putting it inside the model collapses two steps into one and removes a handoff where quality usually degrades.
Speed changes the shape of the work as much as the cost. When a generation takes a minute, a designer writes one careful prompt and waits. When it takes six seconds, the designer writes ten and picks. That difference favors tools that can be called in a loop, which is why the fast variant is likely to end up inside agent workflows in Codex, where a program decides how many images to make and what to do with them.
The two-tier API is also a pricing decision in disguise. Offering a fast, cheap version and a slower, better one lets OpenAI segment customers by urgency rather than by willingness to pay for access. A startup prototyping a landing page wants the six-second model. A retailer producing a catalog wants the one that keeps the product identical across a hundred placements. Charging both the same price would either overcharge the first or lose money on the second.
The competition has moved in the same direction. Image models from rival labs have converged on editing rather than pure generation, on consistency rather than novelty, and on the plumbing that lets a model sit inside an existing production pipeline. The differentiator is less often the most striking sample on a launch page and more often the failure rate on the tenth revision.
There are unresolved questions that no release note answers. Provenance is one: a model that can produce a convincing looping animation of a real product raises the same labeling problem as a still image, and the tools for marking synthetic media are still unevenly deployed. Rights are another. Brands want their own assets preserved, and they want everyone else’s protected, which are two different technical requirements that often get bundled into one policy.
Cost is the third. A complexity parameter is a useful control, but it also shifts the burden of tuning onto the developer, who must now learn what level of fidelity a given task actually needs. That is a real skill, and it is the kind of skill that tends to migrate into tooling: defaults, templates and guardrails that encode somebody else’s hard-won judgment.
For OpenAI the release shows where the competitive pressure now sits. Text models are judged on reasoning. Image models are judged on whether a creative director will accept the output without asking for another round. The second standard is stricter in practice, because it is applied by people with deadlines and no interest in what the model can do when everything is set up in its favor.
Six seconds is a good headline. The question that will decide how widely this model is used is what the picture looks like on the twentieth attempt, when the brief has changed twice and the client wants the same character, a different pose, and the label legible at thumbnail size.
