ChatGPT Images 2.5: Flare vs Sunburst for AI Filmmakers

Published September 16, 2026 — based on the September 8 release of ChatGPT Images 2.5 and our first week of production use at H3 AI Films.

Quick answer: ChatGPT Images 2.5 is OpenAI’s new image generation system, released on September 8, 2026, and it ships as two separate models: GPT-Image-2.5 Flare, the fast default that delivers higher quality than GPT-Image-2 at roughly half the latency, and GPT-Image-2.5 Sunburst, the slower premium model built for campaign creative, product imagery, and precise multi-turn editing. Use Flare for volume and iteration; use Sunburst when a single image has to be perfect. And one thing to be clear about up front: neither model generates video. For AI filmmakers, their real job is generating keyframes that you then animate in a video model like Seedance 2.5 or Gemini Omni 1.1 Flash.

What actually launched on September 8

OpenAI rolled ChatGPT Images 2.5 out across every ChatGPT tier — including Free — on web, desktop, and mobile, alongside the API. The release replaces GPT-Image-2 as the default image system inside ChatGPT and brings a set of workflow features with it: a new Sketch tool for drawing rough compositions the model then renders, reusable templates, prompt sharing, improved non-Latin text rendering, and two generation modes — Instant for speed and Thinking for harder compositions.

Output resolution goes up to 2K. That matters for filmmakers, because a 2K keyframe holds up noticeably better as a first frame for image-to-video generation than the 1K-class outputs older models produced.

If you came here because you heard about “ChatGPT 2.5,” this release is what people mean. There is no new ChatGPT model called 2.5 — the name refers to the image system, and the “two models” are Flare and Sunburst.

GPT-Image-2.5 Flare: the fast default

Flare is what ChatGPT now uses unless you ask for something else. OpenAI’s positioning is simple: better quality than GPT-Image-2 at about 50% lower latency. In practice that makes Flare the volume model — social content, thumbnails, moodboards, and rapid iteration when you are exploring a look.

In our first week of production use, Flare’s speed changed how we storyboard. Exploring eight lighting directions for a scene used to be a coffee-break job; with Flare it happens inside the conversation, fast enough that you stay in creative flow. For a studio, that iteration speed is worth more than any single-image quality gain.

GPT-Image-2.5 Sunburst: the premium finisher

Sunburst is the “most capable” model of the pair. It is slower, and it is built for the images that actually ship: campaign creative, product imagery, brand assets. Its two defining strengths are tighter multi-turn editing — you can push through several rounds of “change this, keep everything else” without the image drifting — and selectable quality settings from low to max.

Multi-turn edit stability is the feature filmmakers should care about most. Keyframe work is iterative by nature: you lock a character, then adjust the pose; you lock a set, then change the time of day. Models that redraw the whole frame on every edit break character consistency. Sunburst holds the frame together far better than GPT-Image-2 did.

Flare vs Sunburst: side by side

No, it is not a video model — and that’s fine

ChatGPT Images 2.5 generates still images only. OpenAI currently has no consumer video model at all: the Sora app closed in April 2026, and the Sora 2 API shuts down on September 24, 2026 — we covered the full story in our Sora vs Veo vs Runway comparison.

So where does an image model fit in a video pipeline? Everywhere, actually. Modern AI video production is largely image-to-video: you generate a precise first frame (and often a last frame), then hand it to a video model to animate. The image model controls composition, character, wardrobe, set, and lighting; the video model controls motion. The better your keyframe, the better your shot.

The keyframe workflow we use at H3 AI Films

Here is the pipeline we run daily, updated for Images 2.5:

Step 1 — Explore with Flare. Generate composition and lighting variants fast. Pick the direction worth investing in.

Step 2 — Finish with Sunburst. Rebuild the chosen frame at max quality, then use multi-turn edits to lock wardrobe, props, and set details without the character drifting.

Step 3 — Animate in a video model. Feed the finished 2K frame into Seedance 2.5 (our workhorse — up to 30-second single shots with native audio) or Gemini Omni 1.1 Flash as the first frame, and direct the motion from there. Our ranked guide to the best AI video tools covers which video model wins which job.

Step 4 — Grade and finish. Upscale, color grade, and cut as usual.

This split — image model for control, video model for motion — is how you get brand-consistent AI video rather than pretty-but-random clips. It is the single biggest quality lever in product and ecommerce video work, where the product has to look exactly right in every frame.

Access and pricing

Both models are rolling out to all ChatGPT tiers, including Free, with usage limits that scale by plan — and both are available in the API as gpt-image-2.5-flare and gpt-image-2.5-sunburst with token-based pricing. For most creators, the practical answer is that Flare is effectively included in whatever ChatGPT plan you already have, and Sunburst is what you reach for when the image matters enough to wait for.

Verdict for AI filmmakers

ChatGPT Images 2.5 will not replace your video model, and it does not try to. What it replaces is the slowest part of the keyframe stage. Flare makes exploration nearly free, and Sunburst’s edit stability solves a genuine production problem — keeping a character or product identical across iterative changes. Pair it with a strong image-to-video model and you have a faster, more controllable pipeline than anything that existed in August.

Frequently asked questions

Is ChatGPT Images 2.5 a video model?

No. Flare and Sunburst generate still images only, up to 2K resolution. For video, pair them with an image-to-video model such as Seedance 2.5, Gemini Omni 1.1 Flash, or Runway — the Images 2.5 output becomes your first frame.

What is the difference between Flare and Sunburst?

Flare is the fast default — higher quality than GPT-Image-2 at roughly half the latency, ideal for iteration and volume. Sunburst is the premium model — slower, with selectable quality up to max and much tighter multi-turn editing, ideal for final campaign imagery and keyframes.

Can I use ChatGPT Images 2.5 for free?

Yes. OpenAI is rolling it out to all ChatGPT tiers including Free, with usage limits that scale by plan. Both models are also available through the API.

What happened to Sora?

OpenAI discontinued Sora 2 in 2026: the app closed on April 26 and the API shuts down on September 24, 2026. As of now OpenAI has no video generation product, which is why its image models are best understood as keyframe generators for other companies’ video models.

Which video model should I pair it with?

For 30-second single-shot spots with native audio, Seedance 2.5. For quick clips with audio inside the Google ecosystem, Gemini Omni 1.1 Flash. For text-to-video benchmark quality, Alibaba Wan 3.0. Our best AI video tools guide compares all of them in production terms.