The product image is approved. Now change the price, keep the bottle recognisable and make it move. That second request is where a striking AI picture starts becoming production work. OpenAI’s Images 2.6, announced on 4 November, brings GIFs and more control over revisions into that gap. The first independent tests also show why a faster generation and an unchanged price list need careful reading.

Based on 8 reviewed sources · See sources

A picture that can move — and return

OpenAI’s launch introduces animated GIFs and saved reference images in ChatGPT, alongside Flare and Sunburst models for developers. References can be recalled in later chats: a useful promise for anyone trying to keep the same character or product across several pieces of work. The company positions Flare for speed and Sunburst for detail and consistency through edits.

The release is global, with no separate European exclusion announced. Its help page adds an access qualification: GIF creation is rolling out. ChatGPT can animate an uploaded still, but cannot edit an uploaded animated GIF. OpenAI’s original announcement shows the intended motion and reference workflow; those demonstrations do not establish production reliability.

From the source · OpenAI ·

OpenAI introduces Images 2.6, GIFs and saved references, with separate developer models.

View the original post on X ↗

The useful control is what stays put

For developers, the image-generation guide describes a way to draw instructions over an image, directing an edit to a particular area. A separate “complexity” setting controls how much planning the model does; quality controls rendering detail. More planning can add time or unwanted changes, so turning everything up is not automatically the safest revision.

The API also offers SVG output: graphics made from editable shapes rather than a flat grid of pixels. That route first generates a reference picture, then uses another model to write the vector file. Both stages are billed. It is a potentially useful hand-off to a designer, with a different cost structure from downloading a still image.

The same token rate can buy a different image

OpenAI’s standard API price remains $30 per million image-output tokens, the units used to meter generated images, matching Images 2.5. Text and image inputs are billed separately. That does not fix the cost of each picture: the number of units can change.

Nitish Garg, founder of AI-agent provider CellCog, reports 80 launch-night generations with matched prompts, sizes and quality settings. In its 1,024-pixel pictogram comparison, Flare 2.6 used many more output tokens at low and medium quality, while high quality used the same count. Flare was faster in that test; OpenAI’s broader claim of 50% faster generation is not a guarantee for every model and request.

The detail

One prompt, a different billable output

CellCog’s matched Flare pictogram test · 1,024 × 1,024 pixels · image-output tokens

One prompt, a different billable output. CellCog’s matched Flare pictogram test · 1,024 × 1,024 pixels · image-output tokens
Quality2.5 → 2.6
Low196 → 824
Medium439 → 1,449
High1,758 → 1,758
One prompt and a commercial provider’s reported runs, not a universal cost multiplier. Counts exclude input tokens. Equal quality labels do not establish equal hidden computation or visual quality; latency observations came from a single launch-night setup.Source: CellCog’s retained launch-night comparison · 2026-11-05

A working GIF is only one part of the hand-off

Curtis Pyke’s separate Kingy test retained prompts and usage for seven API requests: three still-image tasks on each Sunburst version and one native GIF. Its published bottle animation moves, while the revision example changes the requested colour and price. These are small, single-attempt demonstrations. Downloaded website files differ from the hashes in its receipts, so we cannot authenticate them as the original API outputs.

A follow-up CellCog vector test reports a different obstacle: only nine of 48 valid Images 2.6 requests returned an SVG, with server errors on the others. Its nine Recraft requests all returned files; however, the request counts and retries differ. This is a launch-period reliability report, not a general quality ranking or an explanation of the server failures.

Judge the next revision

The release makes a credible new question possible: can a still become a repeatable, editable asset without losing the details someone approved? The documentation offers more ways to attempt it; the first tests expose costs and failure points that a polished launch clip cannot answer.

Our earlier visual editorial, The image is ready. Is the campaign?, follows the work after generation. Images 2.6 changes the tools available at that stage. Whether it shortens the whole job still depends on the revisions, exports and approval work that follow.

Sources

Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.