ChatGPT Images 2.6 now makes animated GIFs and remembers your characters, and AI filmmakers should test it for storyboards

OpenAI has just given its image tool the two things storyboard artists keep asking for: motion and memory. ChatGPT Images 2.6 turns a prompt or an image into an animated GIF, and it can keep your reference images of characters on file so you never upload them twice.

That is a real change for anyone building a look before they commit video credits. You can now sketch a character, loop a gesture, and reuse the same character in the next scene. It is not video generation. But it is a faster way to find out whether a character holds together before you spend on shots.

What OpenAI actually announced

Per @OpenAI, Images 2.6 is available today in ChatGPT and Codex across all tiers, on web, desktop and mobile. The GIF tool lets you adjust speed, edit individual frames and share the result on repeat. A separate feature lets you save reference images of yourself, pets and characters and reuse them in new creations.

For developers, two new models arrive in the API: GPT-Image-2.6 Flare and Sunburst. Flare generates images in as little as six seconds. Sunburst aims for higher quality and greater consistency across edits. Both add image markup and SVG export. OpenAI also says the two models are 50% faster at image generation.

The API is already reachable through third parties. @fal lists GPT Image 2.6 with markup editing, where you draw circles, arrows and notes on an image to guide the edit. It also lists animated GIF output on both text-to-image and edit, and five reasoning effort levels from low to max.

Where it fits in a video workflow

The markup editing is the most interesting part for filmmakers. Circling a hand and writing a note is closer to how a director gives feedback than rewriting a prompt. If it works as described, it should cut the number of rerolls on a hero frame.

The saved references matter for consistency. Most of us already keep a character sheet and re-upload it into every tool. Having it stored in the image tool removes a step, though it only helps if you generate your key frames inside ChatGPT.

@higgsfield_ai has been posting demos made on its platform, mostly small games built from GPT Image 2.6 sprite sheets. One post says the model kept a consistent style across every GIF sprite sheet in a junkyard game. Another says two anime fighters stayed consistent across every edit. These are vendor showcases, so treat them as best cases, not benchmarks. Still, sprite sheets are a harsh test of consistency, because every frame must match the last.

What was not said

OpenAI did not give API pricing for Flare or Sunburst in the posts we saw. The 50% speed claim has no stated baseline. There is no resolution or frame count for the GIFs, and GIF as a format is limited in colour and length, so this is for loops and look tests, not final delivery.

Nothing here makes a video clip. If you need motion at cinematic quality, you still need a video model.

What to test first

- Save one character as a reference and generate the same character in five different scenes. Check face, costume and proportions.
- Use markup to fix one specific flaw in a frame and see whether the rest of the image stays put.
- Loop a short action as a GIF, then compare it with the same idea in your usual video model. It is a cheap motion test.
- Try Flare through fal for quick drafts and Sunburst for the frames you intend to keep.
- Export an SVG of a title card or graphic and check it scales cleanly.

The practical gain is cheaper, faster pre-visualisation with a consistent cast. Test that before you believe anything bigger.

Sources: OpenAI announcement, GIF feature post, Flare and Sunburst API models, availability post, saved references post, fal availability post, Higgsfield consistency demo.