TL;DR
OpenAI shipped GPT-5.6 on July 9, and people are getting it to build finished motion graphics videos from a single prompt. It's not a video model, and it doesn't paint pixels. It writes the animation as code — the same HTML and GSAP that powers slick websites — and that code renders to a clean MP4. That's why the output is precise instead of melty. The catch: the good results come from letting it think for an hour or two on max effort, which costs real money and needs a technical setup. But the quality is the story. This is the moment a general-purpose AI model got good enough to do work that used to need a motion designer.
What actually shipped
GPT-5.6 landed publicly on July 9, 2026, after a limited preview in late June. It's a family, not a single model: Sol is the flagship, Terra the balanced one, Luna the cheap fast one. OpenAI calls Sol its best coding model yet, and there's a max effort setting plus an ultra mode that lets it delegate work to subagents on hard problems.
None of that sounds like video. That's the point, and it's why the story got missed. GPT-5.6 isn't a video model. It's a reasoning and coding model that got good enough at writing code that it can now build an animation, which turns out to be a much more precise way to make motion graphics than generating footage.
So how does a chat model make a video?
It writes it. You describe the video, the model writes the animation as code — HTML, CSS and GSAP, the same web technology behind the slickest sites you've seen. Then that code gets rendered frame by frame into an MP4.
This is the whole reason the output is good. A video generator like Sora guesses at pixels, which is beautiful for a moody establishing shot and hopeless for a logo that has to hold its shape and text that has to stay readable. Code doesn't guess. If you tell it the logo enters at 0.4 seconds and the headline is your exact brand orange, that's what happens, every frame, every time. It's the difference between describing a painting and writing the instructions to build one.
Why does it take hours?
Because the good stuff comes from thinking. On max effort, GPT-5.6 will reason for a long stretch before it writes, and with ultra mode it can spin up subagents to work on parts in parallel. People are letting it run an hour or two on a single clip. It also now inspects the rendered result rather than just firing off code blind, so it catches things that look wrong and fixes them before handing back. That's the loop a motion designer runs, just automated, and it's slow for the same reason good work is slow.
The trade-off is money. Sol is priced at $5 per million input tokens and $30 per million output, and a long reasoning run on a complex animation burns through that. It's not $200 a clip, but it isn't free either, and the DIY path means you're also standing up the render pipeline yourself.
Is the output actually good?
Yes, and that's what makes this worth writing about. Kinetic typography, logo reveals, launch reels, animated charts, product promos — the on-brand graphic motion most marketing teams actually ship. Done well, it's the kind of thing that used to mean an afternoon in After Effects or a week waiting on a freelancer.
Where it still falls down is exactly where you'd expect: cinematic VFX, 3D, hand-crafted character animation, frame-perfect broadcast craft. A senior motion designer in After Effects still wins that, and will for a while. But that was never the work most teams needed. The work most teams needed just got automated. We go deeper on where that line sits in how to make motion graphics without After Effects.
GPT-5.6 vs the video models
| GPT-5.6 (writes animation as code) | Sora / Runway / Veo (generate footage) | |
|---|---|---|
| What it makes | Graphic motion: logos, text, layout | Cinematic, photographic footage |
| Stays on brand | Yes — code is exact | No — it drifts and warps |
| Readable text, intact logo | Yes | Unreliable |
| Time per clip | An hour or two on max effort | Minutes |
| You need | A render setup, or a product built on it | Just the tool |
| Best for | Launch reels, logo reveals, social ads | Mood, b-roll, atmosphere |
What this actually means
A general-purpose chat model, one built for coding and enterprise work, is now good enough to make motion graphics as a side effect of being good at code. Nobody at OpenAI shipped a motion graphics feature. It just fell out of the model getting smarter. That should tell you where this is going: every time the frontier moves, this gets better, for free, without anyone building a video feature at all.
The honest gap is the workflow. Hours of thinking, a render pipeline to wrangle, tokens to pay for, and a fresh round of fiddling every time a scene lands wrong. The capability is here. The convenience isn't, unless someone builds it for you. That's the part worth watching, and it's the part we work on: fluos runs on this same generation of models and turns the two-hour DIY loop into a chat and a live preview. Across the first 500 videos made on fluos, the average render finished in about three minutes. For the wider landscape, our honest roundup of the best AI motion graphics tools compares the options.
Frequently asked questions
- Can GPT-5.6 make videos?
- Not as a video model. It writes animation as code (HTML and GSAP) that renders to an MP4, which is how people are getting finished motion graphics from a single prompt. It doesn't export a video from the chat window itself.
- Why does GPT-5.6 take hours to make a video?
- The good results come from max effort reasoning, where the model thinks at length and can delegate to subagents before writing the animation. Long reasoning runs take time and cost tokens.
- Is GPT-5.6 better than Sora for motion graphics?
- For on-brand graphic motion — logos, text, layout — yes, because code is exact and generated footage drifts. For cinematic photographic footage, Sora is the right tool. They solve different problems.
- How much does it cost?
- GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output. A long reasoning run on a complex animation uses a meaningful amount, so cost scales with how hard you let it think.
- When was GPT-5.6 released?
- Publicly on July 9, 2026, following a limited preview in late June. It comes in three tiers: Sol, Terra and Luna.