Executive summary
On November 4, 2026, OpenAI released Images 2.6 across ChatGPT, Codex and the API. Generation is 50% faster than before, GIF animation, API image markup and SVG export are new, and the model now separates scene planning from pixel rendering through independent complexity and quality parameters. OpenAI says users generate 4 billion images per week worldwide.
What happened
OpenAI launched Images 2.6 on November 4, 2026, simultaneously in ChatGPT, Codex and the API. The announcement frames the release around faster iteration and more precise editing, delivered through four changes: generation speed, GIF animation, expanded API editing controls, and availability.
The scale figure OpenAI publishes is 4 billion images created per week across the three surfaces, averaging more than 500 million per day. At that volume, even a modest per-image improvement compounds into meaningful compute and latency savings, which is why the 50% speed gain is the headline claim.
Unlike earlier releases that focused mainly on output quality, Images 2.6 splits how long the model plans from how finely it renders. For design, e-commerce and marketing teams that edit in batches, that separation determines whether image generation can enter an existing production pipeline rather than remain a one-off experiment.
Key facts
| Item | Detail |
|---|---|
| Release date | November 4, 2026 |
| Publisher | OpenAI |
| Product | Images 2.6 |
| Surfaces | ChatGPT, Codex, API |
| Speed | 50% faster than the previous version |
| Global volume | 4 billion images per week, over 500 million per day (vendor figure) |
| API models | GPT-Image-2.6 Flare, GPT-Image-2.6 Sunburst |
| Fastest output | Flare at quality=low, about six seconds |
| Quality tiers | Sunburst at medium/high for higher fidelity and cross-edit consistency |
| New content capability | GIF animation, sprite sheets |
| New API capabilities | Image markup, SVG export, complexity parameter |
| Availability | Global; all ChatGPT and Codex subscription tiers, desktop, mobile and web |
| Provenance | C2PA content credentials and Google DeepMind SynthID |
Speed: 50% faster and six-second images
OpenAI says Images 2.6 generates images 50% faster than its predecessor. In the API, speed and quality are split across two models: Flare targets speed and layout, Sunburst targets detail and consistency.
Flare is built for throughput. With quality set to low, it can generate an image in as little as six seconds. OpenAI also says Flare improves layout quality for posters, product imagery and other professional visuals. Sunburst is built for stability: at medium or high quality it delivers higher fidelity and better preserves characters, products and brand details across editing rounds.
The two tiers map onto two real workflows. Social and e-commerce teams chasing volume benefit from a six-second loop that makes generate-select-regenerate feel nearly instant. Brand key visuals and projects that need repeated revision care more about Sunburst's consistency across edits than about how striking the first frame looks.
GIFs built from sprite sheets
Images 2.6 adds GIF generation. OpenAI lists the use cases as animating a brand mascot or character, turning a product shot into a looping visual, making reactions and memes for social media, and exploring character movement with animated storyboards before video production.
The mechanism does not call a video model. It generates a 3-column by 2-row sprite sheet with six sequential frames inside a single image, then assembles them into a seamless looping GIF. The published example prompt requires exactly six sequential frames in a precise 3x2 grid read left-to-right then top-to-bottom, identical panel dimensions, a fixed camera, the same character and background, and explicitly forbids captions, numbers, logos, speech bubbles and borders. The constraint pushes continuity into layout rather than frame interpolation.
For creators, the gain is cost. A looping animation previously required either a video model or frame-by-frame generation plus alignment; now a single image request yields usable six-frame material. The limit is equally clear: six frames can carry only so much action and emotion, and longer narratives still need a video model.
API: separating planning from rendering
The three API changes are aimed at engineering workflows.
Image markup comes to the API. The capability already existed in ChatGPT; developers can now specify exactly where an edit should happen instead of describing the whole image in text. For batch retouching, local changes become structured instructions.
SVG export. Generated output can be exported as vectors for websites, apps and print, removing a redraw step for logos, icons and layout work.
A complexity parameter. OpenAI describes it as how much planning the model does to satisfy a request, while quality controls visual detail and fidelity. For most editing tasks the company recommends low complexity for faster edits paired with whatever quality the workflow needs; scenes with many elements or constraints should use medium or high.
The split changes the cost model. Changing a color no longer requires re-planning the entire scene, so compute can be spent on rendering alone. OpenAI also notes that complexity and quality are independent dimensions; raising one does not automatically raise the other.
Safety evaluation and provenance
OpenAI published a system card for Images 2.6. The evaluation used 3,077 adversarial prompts spanning policy violations and high-risk categories including biological and cyber content.
According to the card, Sunburst's violative output rate is level with or slightly below the 2.5 generation. Flare's rate rose from 1.49% (46 of 3,077) to 2.14% (66 of 3,077), an increase of 0.65 percentage points that OpenAI describes as not statistically significant. The models did not cross the high-risk threshold for biological or cyber capability, but OpenAI still applies high-biological-risk mitigations. GIF generation has a separate safety stack because animation can carry dynamic content that static frames would not trigger.
On provenance, all Images 2.6 outputs across ChatGPT, Codex and the API carry C2PA content credentials and Google DeepMind's SynthID watermark. That gives platforms and brands a technical handle for separating generated images from photographs, but it proves only which model produced an image, not that the image is true.
Availability and industry impact
Images 2.6 launched globally the same day. ChatGPT and Codex cover every subscription tier on desktop, mobile and web. The API offers Sunburst and Flare, billed by tokens with images converted to output tokens; the announcement does not publish a separate price table, so real cost still depends on resolution, quality tier and image count.
Three effects stand out. First, the speed floor for image generation keeps dropping, and six-second output makes live previews and batch drafts routine. Second, GIFs and sprite sheets pull lightweight motion back from video models, shortening the marketing production chain. Third, decoupling complexity from quality makes generation look more like a tunable industrial process and less like a slot machine.
The limits are just as real. A six-frame structure cannot carry long narratives; SVG export solves format, not node-level editability; and improved multi-turn consistency is still largely a vendor claim that needs validation on specific brand assets. Teams wiring generation into production should first run a controlled comparison on the same brand materials across Flare versus Sunburst and low versus high complexity, then set defaults.
Official sources
- OpenAI announcement: Introducing Images 2.6
- OpenAI Japanese edition (same content): Images 2.6 のご紹介