What people are building with OpenAI's new model, the prompts behind it, and the plugin that lets it generate any image or video it needs while it builds.
The plan · 6 steps
- 1Switch to Work and pick Astra
- 2Start the Higgsfield trial
- 3Add the plugin and authorize
- 4Turn Higgsfield on in the chat
- 5Run the motion website prompt
- 6Name the output, ask for a preview
- What GPT-6 Astra actually is, in plain English, and where to find it
- The builds from the video: a 2,000-piece human body, a walkable house, a game from one sentence
- Twelve more use cases, with the prompt shape that gets each one
- Connecting the Higgsfield plugin inside ChatGPT, click by click
- Get Higgsfield with a free trial (link inside)
- The motion-website prompt from the video, expanded so it works first time
- Astra plus Higgsfield: UGC ads, explainers, game art, and the credit math
What Astra Actually Is
OpenAI shipped GPT-6 Astra in early September and called it "the world's most intelligent and aligned model." Strip the marketing and here's what changed for you: it doesn't just answer, it builds. It uses a computer and a browser, writes and runs real software, models in 3D, makes games with graphics and physics, and publishes websites straight from the chat.
It's rolling out to ChatGPT Plus, Pro, Business and Enterprise, with a heavier Astra Pro tier for Pro and Business. You'll find it in the model picker, and in Work mode (the switch at the top of ChatGPT) it gets the tools that make the wild stuff possible: a browser, code execution, file building and ChatGPT Sites for hosting.
- 99.9% on ARC-AGI-3, per OpenAI
- 1.9x faster at web tasks than the last model
- 1 sentence to a playable game
The one-line version. Astra is the first ChatGPT model where "build me X" reliably ends with X existing: a file, a site, a model, a game. That's why the demos look unhinged. Same prompts as before, wildly different finish line.
Where the use cases in this guide run
| Type of build | Where it happens | What you get |
|---|---|---|
| 3D models, Blender, CAD | Work mode, code execution, a Blender or CAD file it writes for you | A file you open in Blender, or a viewer in the chat |
| Games | Work mode writes the code, ChatGPT Sites hosts it | A link you can play in a browser |
| Websites | Work mode plus ChatGPT Sites | A live URL, hosted by OpenAI |
| Images and video | The Higgsfield plugin inside ChatGPT (chapter 4) | Generated media it uses inside the build |
| Documents, decks, spreadsheets | Work mode | Finished files |
Tip. Everything in this guide assumes Work mode is on. If Astra is answering in paragraphs instead of building, you're in Chat.
The Builds From The Video
These are the four that made me stop scrolling. Each one comes with the shape of prompt that produced it.
1. A human body that splits into 2,000 pieces
Someone asked Astra for a full 3D model of the human body where every part is a separate, labeled piece, then an exploded view. It wrote the geometry, built the hierarchy, and exported a file that opens in Blender with 2,000 selectable parts.
PROMPT SHAPE: THE 3D MODEL
Build a 3D model of [SUBJECT] where every component is a separate named object organized in a hierarchy ([system > organ > part]). Export as a .blend or .glb I can open in Blender, plus an "exploded view" where the parts separate along their axes on a slider. Show me a rendered preview before you export.
2. A house you can walk through
A full house modeled in Blender from a floor plan and a few reference photos, then handed to a game engine so you can walk it in first person. OpenAI showed the same trick in their own launch demo, house to Unreal Engine walkthrough.
PROMPT SHAPE: THE WALKTHROUGH
Here's my floor plan [attach] and photos of the style I want [attach]. Model the whole house in Blender with real dimensions, materials and lighting, then build a first-person walkthrough I can open in a browser (WASD to move, mouse to look). Host it and give me the link.
3. The Mona Lisa, painted inside Photoshop
This one's the Higgsfield crossover. Astra drove Photoshop through the browser and used Higgsfield to generate the strokes and reference layers, rebuilding a famous painting layer by layer while the person watched. The lesson isn't the painting, it's that Astra can operate a real desktop app and pull generated images into it mid-task.
4. A playable racing game from one sentence
PROMPT SHAPE: THE GAME
Build me a playable browser game: [one sentence, e.g. a low-poly car drifting around a coastal track at sunset with a lap timer]. Keyboard and touch controls, single file, host it on ChatGPT Sites and give me the link. Name it [NAME].
Twelve More Things People Are Building
Pulled from OpenAI's launch demos and what's been circulating since. Each row is a real category of build with the prompt opener that gets it started.
| Build | Say this to start |
|---|---|
| Circuit board layout | "Design a PCB in KiCad for [device]: schematic, board layout, and a parts list I can order." OpenAI did one in under three minutes. |
| CAD from photos | "Here are photos of [object] from four angles. Reconstruct it as a CAD model with real dimensions and export STL for 3D printing." |
| A full app, hosted | "Build and publish a web app that [does X], with login, a database, and a mobile layout. Show me a preview, then publish." |
| Slideshow from a doc | "Turn this report into a 12-slide deck in my brand colors with one chart per section." |
| Spreadsheet model | "Build a spreadsheet that forecasts [X] with inputs I can change on the first tab and a chart that updates." |
| Legal formatting | "Reformat this contract to match [firm's] style guide, fix numbering and cross-references, keep every word." |
| Browser agent | "Open [site], find [thing], compare the top five, and fill my cart with the best value. Ask before you pay." |
| Research report | "Read these twenty papers and give me a literature review with a table of methods and results, cited." |
| Interactive lesson | "Build a browser lesson that teaches [topic] with a quiz at the end and a score I can share." |
| Data dashboard | "Connect to this CSV and build a live dashboard with filters, hosted, that I can send to my team." |
| Music visualizer | "Make a browser page that reacts to my mic input with 3D shapes, in my brand colors." |
| Physics sandbox | "Build a 2D physics playground where I can drop shapes, change gravity, and record a GIF." |
Tip. The pattern in every one: name the output (a file, a link, a deck), name where it lives (Sites, Blender, my desktop), and ask for a preview before publish. Astra fills in everything else.
Connect Higgsfield Inside ChatGPT
Astra can build almost anything, but it can't generate cinematic video or high-end images on its own. That's what the Higgsfield plugin adds: Seedance, Kling, Nano Banana Pro, GPT Image 2 and Higgsfield's own skills, all callable from the chat. Astra plans, Higgsfield renders.
Get Higgsfield (free trial)
Start here: higgsfield.ai, free trial for new users. You need a Higgsfield account to connect the plugin, and generations spend Higgsfield credits, so start the trial before you connect. Cancel the auto-renewal inside Higgsfield if you're only testing.
Step 1: Add the plugin
Go to higgsfield.ai/mcp and click Add Higgsfield plugin, or in ChatGPT open the plugins directory, search Higgsfield, and click Add. This is the official listing, no MCP link to paste.
Step 2: Sign in and authorize
A Higgsfield window opens. Sign in to your Higgsfield account and click authorize. That's the whole connection: no API key, no settings.
Step 3: Turn it on in the chat
In a new chat, switch to Work and pick Astra. Open the tools menu by the message box and make sure Higgsfield is on. Test it: "Use Higgsfield to generate a 16:9 image of a sunrise over a desert highway." If an image comes back, you're set.
What works in ChatGPT, and what doesn't
Inside ChatGPT you get image generation, video generation and the content skills (Ad Multiplier, Editorial Motion Graphics, Product Review UGC, Faceless Content, Stickman Cartoon). Two things stay on Higgsfield's own site: audio and voice, and Higgsfield's own website-building skill. So in the video, Astra built the site with ChatGPT Sites and Higgsfield made the motion video. That's the division of labor to expect.
Heads up. Troubleshooting. "Authorization failed": in ChatGPT, Settings, then Plugins, then Higgsfield, reconnect. Images work but video fails: your Higgsfield plan or trial is out of credits. Video is the expensive part; every generation deducts credits whatever plan you're on.
The Motion Website Prompt
This is the prompt from the video, expanded. I pasted the short version and Astra rebuilt my site with my fonts, colors and images and put the Higgsfield video right on the front page. The long version below gets that result first time instead of after three clarifying questions.
PROMPT: BUILD MY MOTION WEBSITE
Rebuild my website as a motion website. My current site is [URL]. First pull my fonts, colors, logo and the images already on it so the new build matches my brand exactly. Then use Higgsfield to generate a silent, looping 8-second hero video for the top of the homepage: [the vibe: e.g. warm desk-lamp light, my product in frame, slow push-in, no people, no text]. Use the same still as the start and end frame so it loops cleanly. Build the whole site in ChatGPT Sites: home with the video hero (muted autoplay), scroll-reveal sections, an about page, a page for my [products/guides], and a contact section that sends to [EMAIL or form link]. Keep my existing text where it's good, rewrite what's weak and show me each rewrite next to the original. Don't invent facts about me. Show me a preview before you publish, then publish it and give me the link.
Why it's worded like that
Silent because audio isn't available through the plugin in ChatGPT, and a muted autoplay hero is what every modern site does anyway. Same still as start and end frame is the only reliable way to get a seamless loop; without it the video jumps every eight seconds. Don't invent facts about me keeps it from writing you a bio you never had. Preview before publish is your last click.
Astra Plus Higgsfield: The Combined Use Cases
Astra does the thinking, structure and code; Higgsfield does the pixels. Three prompts I tested.
UGC REVIEW + FIVE AD VARIATIONS
Here's my product photo [attach]. Use Higgsfield's Product Review UGC skill to make three 10-second 9:16 clips of someone showing it (pick up, close-up on [FEATURE], hold to camera), silent with on-screen text, plus a stitched cut. Generate one clip first for my approval, then the rest. Then use Ad Multiplier to make five 15-second Reels variations with different hooks and a caption for each. Only claims from my product page: [URL].
BLOG POST TO FACELESS EXPLAINER
Turn this post [paste] into four 15-second 9:16 scenes with Higgsfield's Editorial Motion Graphics skill. On-screen text quoted from the post, under 8 words per line, plain sans-serif, colors [BG HEX] and [ACCENT HEX]. Storyboard as a table first, mark what you trimmed, then generate the four clips plus a stitched 60-second cut. No voiceover.
GAME WITH GENERATED ART
Build the racing game from the one-sentence prompt, then use Higgsfield for the loading-screen key art (Seedream, 16:9) and a silent 5-second title animation from that art (Google Omni Flash). Upload both as site assets, don't hot-link, and give me the link.
| Combined use | Astra does | Higgsfield does |
|---|---|---|
| Motion website | Scrapes brand, writes and hosts the site | The hero video |
| Product ads | Captions, hooks, claims check against your page | UGC clips and the five variants |
| Explainers | Storyboard, text trimming, stitching | The four motion-graphics scenes |
| Games and apps | All the code and hosting | Key art, title animation, textures |
The Credit Math And The Mistakes
Video is where the credits go. Images are cheap, a 10-second clip is not, and regenerations are the hidden cost. Here's how to keep a free trial from evaporating on your first prompt.
| Mistake | Fix |
|---|---|
| Asking for a 30- or 60-second video | Clips run about 5 to 15 seconds. Ask for clip count, not total length, and a stitched cut. |
| Expecting a voice | Audio isn't in the plugin. Say "silent" up front, add voice or music in Higgsfield or Instagram after. |
| Batch-generating before checking one | "Generate one clip first for my approval" saves more credits than any other line. |
| Long on-screen text | AI video warps long text. Under 8 words a line, plain sans-serif, or you'll regenerate. |
| Using the premium model for everything | Seedream 5 Lite and Google Omni Flash for drafts and game art; Kling 3 and Seedance 2.5 for the hero shot. |
| Hot-linking Higgsfield files on your site | Upload them as site assets. Links can expire and your hero goes blank. |
| Skipping Work mode | Chat mode explains. Work mode builds. Every prompt in this guide needs Work. |
| No preview gate | Add "show me a preview before you publish." Astra will happily ship version one. |
The one-page version
- ChatGPT. Plus or higher. Switch to Work. Pick Astra.
- Higgsfield. Create the account, start the trial.
- Plugin. higgsfield.ai/mcp or ChatGPT plugins: add Higgsfield, sign in, authorize.
- New chat. Tools menu, Higgsfield on. Test with one image.
- Same chat. Paste the motion-website prompt with your URL and vibe. Preview, then publish.
- Anything else. Name the output, name where it lives, ask for a preview. Let it build.
Do This Today
Start the Higgsfield trial at this link, add the plugin, and run the motion-website prompt on your own site. Twenty minutes, and you'll have a version of your homepage with a video hero that most agencies would have quoted you a month for.
Astra builds it. Higgsfield paints it. You just describe it.