AI Automation

OpenAI Images 2.6: GIFs, reference images and API pricing

By Yogesh · November 10, 2026 · 7 min read

Sources checked and dated on publication. See our editorial policy.

OpenAI released Images 2.6 on November 4, 2026. The new image model generates about 50% faster, adds native GIF output and saved reference images in ChatGPT, and ships two API variants — Flare for speed, Sunburst for edit precision.

The short answer

50% faster generation — OpenAI says Images 2.6 generates images 50% faster than the previous version, and the Flare API variant can return an image in as little as six seconds when quality is set to low.
Two API models — GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst are available in the API, each priced at $8.00 per million image output tokens and $5.00 per million text input tokens.
GIFs and reference images — ChatGPT gained looping GIF creation with adjustable playback speed and saved reference images, so a photo can be reused in later chats without uploading it again.
4 billion images a week — OpenAI says people now create 4 billion images per week across ChatGPT, Codex and the API — more than half a billion per day on average.

What shipped and when

OpenAI's ChatGPT release notes for November 4, 2026 list the user-facing changes plainly:

GIF creation rolling out in ChatGPT — ask for a looping animation, or upload a sketch and ask for the character to wave. Playback speed is adjustable and the GIF is shareable.
Saved reference images let you reuse a photo in later chats. The notes give the example of saving a photo of your dog as "Pepper", then asking for a watercolor portrait of @Pepper without uploading the photo again.
A faster and more consistent image model underneath, described as "better at preserving details across edits, like a character's outfit when you change the background".

OpenAI's own product post adds the scale figure: "People around the world now create 4 billion images per week across ChatGPT, Codex and the API. On average, that's more than half a billion images created per day." That is the number that explains why OpenAI is spending its effort on iteration speed rather than on a new aesthetic look.

Availability is broad: Images 2.6 is available globally in ChatGPT and Codex on desktop, mobile and web, and in the API as gpt-image-2.6-flare and gpt-image-2.6-sunburst.

Flare versus Sunburst: pick your variant

The API exposes two variants of the same generation family, and the split is about control rather than capability tier.

GPT-Image-2.6 Flare — built for speed. When quality is set to low, OpenAI says Flare can generate images in as little as six seconds. Flare also improves layout quality for posters, product imagery and other professional visuals.
GPT-Image-2.6 Sunburst — built for detail and consistency. At medium or high quality it brings "higher-quality image generation and even greater image consistency across edits".

Both are billed identically on the published price list: $8.00 per million image output tokens and $5.00 per million text input tokens for standard processing, with cached input at $1.25. The text input price is not the text you type into the chat box — it is the image input tokens charged when you edit an existing image. That is the line that matters for edit-heavy workflows, because every iteration re-uploads the source image.

The practical read: if your workflow generates many variations from scratch, Flare at low quality is the cheap fast path. If your workflow edits one asset repeatedly and the asset has to stay recognisable, Sunburst at medium or high quality is what preserves the identity across rounds. OpenAI's own guidance on the product post: use low complexity for faster edits, and move to medium or high complexity "for scenes with many elements or constraints".

The three API features that change the workflow

Beyond the model swap, OpenAI added three API capabilities that alter how you structure an image pipeline.

Image markup — previously a ChatGPT-only feature — now comes to the API, so teams can "pinpoint exactly where to make specific changes" instead of describing the region in prose.
SVG export — "make designs ready to use across websites, apps, and print". This is the one that changes downstream workflows, because vector output can be edited and scaled rather than re-rasterised.
A complexity parameter — "lets developers adjust how much planning the model does to meet a request, with lower settings suited to quick edits".

There is also a quality control that matters more than it sounds: the image quality setting controls visual detail and fidelity, while complexity controls how much effort the model spends building the scene. They are separate knobs, and conflating them is how you end up paying for expensive generations that produce simple results.

What it costs in practice

Cost is driven by two variables you control: the quality setting, and how many times you edit. The pricing guide notes that GPT Image models prior to gpt-image-2 bill proportionally to the number of tokens required to render an image, and that larger dimensions and higher quality produce more tokens. The same logic carries into the 2.6 family, where image output is billed per million tokens.

Three habits keep the bill predictable on this API:

Batch where the workflow allows it. Batch pricing sits below standard for image generation models, and a design-exploration job does not need a response in six seconds.
Separate generation from editing. Editing re-reads the source image and charges image input tokens on every round. Generate more candidates with Flare, then commit to one and edit it with Sunburst, rather than iterating on a single expensive output.
Use the low-quality path for exploration. Flare at low quality exists precisely so that the iteration loop does not run at production settings.

For a team already generating images in ChatGPT, the migration question is different: the product post says Images 2.6 is available in ChatGPT and Codex across all subscription tiers, so the upgrade arrives without a pricing decision. The cost question only starts at the API boundary.

Where the GIF feature actually earns its keep

GIF output is the feature with the clearest downstream use, and OpenAI's launch post is careful about how it frames the mechanism. The published example is a sprite sheet — six sequential frames laid out in a 3-column by 2-row grid, read left to right then top to bottom, with every panel the same dimensions, a fixed camera, and the final panel transitioning smoothly back to the first. OpenArt's creative director called GIFs "the native language of social media" in the launch post, and the OpenArt examples are mascot animations and market-chart reaction loops built from single prompts.

The production pattern that follows from this is clear: generate a looping sprite sheet, then convert it to a GIF. The prompt format OpenAI published is unusually prescriptive — exact frame count, exact grid layout, fixed camera, identical character, identical background, one clearly described incremental motion, and a closing frame that loops back to the first. That is a reusable template, and the mention of "animated storyboards to explore character movement before video production" is the honest framing: this is a pre-production tool, not a video generator.

Two named early customers are worth noting for different reasons. OpenArt used it for creator-facing social loops. Town, the startup behind an AI work assistant, used Images 2.6 in the API to explore animated characters for its work assistant — Monty, an animated character that appears above the message box as a typing indicator. Figma's creative director described adding the model to Figma Weave "inside a reusable workflow", and Picsart's GTM director cited "more accurate results, better character consistency, and faster generations" for social campaign work.

Who should use it, and what to check first

Images 2.6 is a strong fit if you are shipping product imagery, social assets or character-led visuals where identity must survive editing. It is a poor fit if your pipeline assumes a raster-only, static-image world and you have not budgeted for SVG output changes downstream.

Verdict Adopt Images 2.6 if you iterate on visuals repeatedly and need character or product identity to hold across edits. Sunburst at medium or high quality is the variant for that, with the complexity parameter raised only when a scene genuinely has many elements.

Use Flare for exploration — low quality, six-second returns — and reserve the premium settings for the assets you actually ship. The 50% speed improvement matters, but only if you also stop paying production quality for drafts.

A one-hour test plan

Pick one asset you already produce: a product shot, a mascot, a blog illustration. Generate ten candidates in Flare at low quality and note the wall-clock time and failure modes. Take the best one and run five Sunburst edits at medium quality, changing background, lighting and framing, and measure how often identity holds. Then export one result as SVG and check whether your downstream tools accept vector output. That single test tells you more about Images 2.6 than any benchmark — it measures your workflow, not a demo.

Frequently asked questions

50% faster generation

OpenAI says Images 2.6 generates images 50% faster than the previous version, and the Flare API variant can return an image in as little as six seconds when quality is set to low.

Two API models

GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst are available in the API, each priced at $8.00 per million image output tokens and $5.00 per million text input tokens.

GIFs and reference images

ChatGPT gained looping GIF creation with adjustable playback speed and saved reference images, so a photo can be reused in later chats without uploading it again.

4 billion images a week

OpenAI says people now create 4 billion images per week across ChatGPT, Codex and the API — more than half a billion per day on average.

Sources

Prices and benchmark figures are standard published rates as of November 10, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.