TL;DR
Fluos is built on the same frontier models behind ChatGPT, and we're not shy about it. The reason is simple: motion graphics turned out to be a coding problem, not a video problem. The best video from AI right now doesn't come from a video model — it comes from a reasoning model writing animation as code. OpenAI just shipped its best coding model yet. That's our engine. What we build is everything between a two-hour DIY render loop and a finished on-brand MP4, which is most of the work. And when the models jump, your videos get better without you doing anything.
What Fluos actually runs on
You describe a video in a chat, an AI writes the animation as code (HTML and GSAP), you watch it build in a live preview, and you export an MP4 or GIF. The AI doing the writing is the frontier generation from OpenAI — the same models behind ChatGPT. We don't train our own video model, and we're not trying to.
That's a deliberate bet, and it's worth explaining, because the instinct in this industry is to hide it.
Why build on someone else's model?
Because motion graphics turned out to be a coding problem.
The obvious way to make AI video is a video model: train something to generate footage, prompt it, get pixels. That's Sora, Runway, Veo. It's genuinely great for mood and b-roll, and it's the wrong tool for a logo reveal. Generated footage drifts. Text warps. Your logo doesn't survive the frame. For anything that has to be exactly on brand, guessing at pixels is a losing approach.
Writing the animation as code isn't. Code is exact. The logo enters at 0.4 seconds because you said so. The headline is your brand orange — that hex — every frame. It re-renders at any size without falling apart. So the question stops being "who has the best video model" and becomes "who has the best coding model." And that's a question OpenAI has been answering, at enormous expense, for years. GPT-5.6 is described as their best coding model yet. We'd be insane to try to beat that instead of building on it.
So isn't Fluos just a wrapper?
Fair question, so here's the honest answer. The model is the engine. Nobody buys an engine.
You can go get GPT-5.6 to write you an animation right now. People do, and the results can be excellent. What they don't tell you is what it takes: standing up a render pipeline, letting the model reason for an hour or two on max effort, paying for the tokens, then rendering, compressing, and going back around when scene two lands wrong. It works. It's not a workflow you run every Wednesday when the launch post needs a video.
What sits between that and a finished video is the actual product: the brand system that keeps your colors and fonts locked across every output, the live preview so you're reacting to the real thing instead of waiting on a render, the surgical edit so you can fix one scene without redoing the whole thing, the aspect ratio variants from one brief, the render pipeline you never see. Across the first 500 videos made on Fluos, the average render finished in about three minutes. That's the gap between the engine and the car.
The DIY loop vs Fluos
| ChatGPT + your own setup | Fluos | |
|---|---|---|
| The engine | Frontier OpenAI model | Frontier OpenAI model |
| Render pipeline | You build it | Handled |
| Live preview | No | Yes |
| Brand locked across outputs | Re-prompt it every time | Built in |
| Fix one scene | Regenerate and hope | Ask for that scene |
| Time to a finished clip | An hour or two, plus fiddling | Minutes |
| You need to be technical | Yes | No |
What happens when the models get better?
Your videos get better. That's the part worth sitting with.
A tool built on templates gets better when someone draws more templates. A tool built on its own video model gets better when that team ships a new one, which is slow and expensive. Fluos gets better every time the frontier moves, and the frontier is moving faster than any single company can ship features. GPT-5.6 landed in July and it reasons harder, writes better code, and now inspects what it rendered instead of firing blind. We didn't build any of that. Our users got it anyway.
The flip side, honestly: we don't control the engine. If OpenAI has a bad day, we feel it. That's the trade, and we think it's obviously the right one, because the alternative is betting we can out-research the labs. We'd rather spend our time on the thing they're not building, which is the workflow between a prompt and a video you'd actually publish.
Why we think we're early, not late
The capability arrived before the convenience. Right now, making great motion graphics with AI means being technical enough to wire it up and patient enough to wait two hours. That's the same shape every technology takes right before it goes mainstream: it works, it's just annoying. The annoying part is the product opportunity, and it's what we've spent our time on while everyone else was watching the video models. For the wider picture, see our roundup of the best AI motion graphics tools, or how to make motion graphics without After Effects.
Go make something and see. fluos.io
Frequently asked questions
- What AI model does Fluos use?
- Fluos runs on the frontier generation of OpenAI models, the same ones behind ChatGPT. The model writes the animation as code, which is what keeps the output precise and on brand.
- Is Fluos just a ChatGPT wrapper?
- The model is the engine. Fluos is the brand system, live preview, scene-level editing, aspect ratio variants, and render pipeline around it. You can prompt ChatGPT for animation code yourself — it just takes a technical setup and an hour or two per clip.
- Why doesn't Fluos train its own video model?
- Because on-brand motion graphics is a coding problem, not a footage problem. Code is exact; generated footage drifts. The best coding models come from the frontier labs, so we build on them.
- Does Fluos get better when ChatGPT gets better?
- Yes. Because the animation is written by a frontier model, improvements to that model show up in the output without users changing anything.