In the days after GPT-6 Astra launched, my feed filled up with 3D demos: exploded product views in the browser, Blender scenes, Unreal Engine environments and small browser games. In the video I go through 20 of them. This article groups them by type, explains what they suggest for real client work, and shows how to try a small version yourself with the access you already have.

One caution up front. Many demos are shared as “one prompt” results. I can’t verify how much iteration, editing or manual cleanup went into each one, so treat them as a picture of what’s possible on a good day, not what you’ll get on your first try.

Prefer to watch? See the full walkthrough on YouTube. I go through every demo on screen.

Who can use GPT-6 Astra right now

In the video I say Astra is available to everyone. That’s not quite right, so here’s the access picture as of late October 2026, based on OpenAI’s own pages:

If you use the Codex CLI, update to version 0.153.0 or later to get Astra.

The kinds of demos people are building

Exploded 3D product views in the browser

The most common demo is an interactive 3D object built with three.js. You drag a slider and the product separates into labeled parts. Hardware, vehicles and mechanical objects are popular subjects. For a product business, this is basically an interactive explainer that runs in any browser without a plugin.

Blender and Unreal Engine scenes

Astra can write scripts that drive tools like Blender, so people are generating product renders, interiors and landmark buildings. A few demos go further into Unreal Engine environments. The model writes the scene logic, and the 3D tool does the rendering.

Real estate and architecture

Several demos turn photos or plans into rough 3D models of homes and villas. In the video I point out that this is where I see the clearest business use: real estate developers and architects who want a quick walkthrough for a client before anything is built.

Browser games

Racing games, block-building worlds and simple shooters built with three.js show up a lot. Most are small prototypes, but they’re playable, and that’s a big step from where coding models were a year ago.

Interface and design work

Some demos recreate desktop-style user interfaces in the browser. Others use computer use to copy a design into a tool like Canva. These are rougher, but they show the model can work across both code and visual tools.

I click through several of these on screen in the video, which gives you a better sense of how smooth they are than screenshots do.

What this means for real client work

Demos are optimized to look impressive in a 30-second clip. Client work has different requirements: accuracy, performance on a phone, edits that don’t break everything, and assets you’re allowed to use. Here’s where I’d use Astra today:

• Product explainers. An exploded view of a product on a landing page or in a sales deck.
• Property previews. A rough 3D walkthrough so a client can react to a layout before you pay for detailed renders.
• Pitch prototypes. A small interactive demo that shows an idea in a meeting instead of a slide.
• Training material. Interactive diagrams of equipment or processes for onboarding.

In each case, plan for a human to review and fix the output. The model gets you a first draft fast. You still own the quality.

If you want structured lessons for turning these experiments into client work, AI Income Lab has Codex tutorials and courses to follow.

How to try your own 3D demo, step by step

1. Pick one small object. Don’t start with a city or a full game. One product with six to ten parts is enough.
2. Choose where to run it. Use Codex if you want files on your machine that you can edit. Use ChatGPT Work if you just want to see a result.
3. Write a specific prompt. Name the library (three.js), the output (one HTML file), the interactions and the constraints. There’s an example below.
4. Run it locally. Open the file in a browser, or serve the folder with a simple local server such as python3 -m http.server.
5. Give targeted feedback. “The GPU label overlaps the RAM, move it 20% right” works better than “make it better.”
6. Keep reasoning effort moderate. Medium handles most of this. Move to high only for a bug the model can’t fix on the first try.
7. Test on a phone. A scene that runs well on a laptop can stutter on mobile. Cut polygon counts and lights if it does.

Example prompt

This is a starter prompt you can adapt. It isn’t taken from any demo in the video.

Build a single-file web page (index.html) using three.js
from a CDN.
Scene: an exploded view of a desktop computer. Show the case,
motherboard, CPU, cooler, RAM, GPU and power supply as simple
low-poly shapes.
Behavior:
- A slider moves the parts between assembled and exploded.
- Hovering a part shows its name in a label.
- Orbit controls for rotate and zoom.
- Soft lighting and a neutral background.
Constraints:
- No external models or textures. Shapes only.
- Must run at 60 fps on a mid-range laptop.
- Comment the code so I can edit part positions later.

For Blender, ask for a Python script that uses Blender’s bpy module, then run it from Blender’s scripting tab. Keep each script to one task, like building the geometry, then the materials, then the camera.

Stuck on the local setup or the Codex CLI? Ask inside the AI Income Lab community, where members work through builds like this.

Common mistakes to avoid

• Expecting a one-shot result. Plan for several rounds of feedback, even when the demo you saw looked instant.
• Starting too big. Large scenes burn through usage and are hard to debug. Build small, then combine.
• Running everything at maximum effort. It’s slow and expensive. Medium is enough for most 3D scaffolding.
• Using brand names and copyrighted designs in client work. A demo of a famous car or game is fine for learning. For a client, use their own products or original designs.
• Skipping performance checks. Many browser demos are built and recorded on powerful machines. Your client’s visitors may not have one.
• Upgrading before you’ve tested. Try Plus with Work and Codex first. Move to a Pro plan only if you keep hitting the limit.

FAQ

Can GPT-6 Astra really build a 3D game from one prompt?

Some demos claim that. I can’t confirm how much iteration went into them. Expect to go through several rounds of feedback for anything beyond a simple scene.

Do I need a Pro plan to try this?

No. Plus includes Astra in ChatGPT Work and Codex with limited usage. Pro plans give you more usage and GPT-6 Pro in regular Chat.

What do I need installed?

For three.js demos, just a browser and, ideally, a simple local server. For Codex, install the CLI (version 0.153.0 or later). For Blender scenes, install Blender itself.

Can I use this for client projects?

Yes, for prototypes and explainers, as long as a person reviews the output and you only use assets you have the rights to.

How much does Astra cost through the API?

OpenAI lists gpt-6-astra at $10 per million input tokens and $50 per million output tokens. Check the pricing page for current rates before you build on it.

Conclusion

The GPT-6 Astra demos are a real step forward for 3D in the browser, Blender scripting and quick prototypes. They’re also curated highlights. The practical move is to start with one small object, run it locally, give specific feedback, and test it on a phone before you show it to a client.

Your next step: copy the example prompt, swap in a product you know well, and see how far you get in three rounds of feedback.

If you’d rather not figure this out alone, join AI Income Lab for the courses, templates and help that go with these builds.