What happened
Google has opened the door to Gemini 4 Carbon, the most capable model it has released so far, and it's no longer locked behind a waitlist. Developers, enterprises and Google AI Ultra subscribers can use it right now through three routes: the API, the Gemini app, and Antigravity — Google's workspace for building with AI agents. (The API is simply the plumbing that lets other software call the model directly.)
But the model itself isn't the headline. It's what Antigravity can now do with it.
/swarm: one request, a whole crew of agents
Antigravity has shipped a preview of a mode called `/swarm`. Instead of walking one chatbot through a long task step by step, you send a single request and a group of agents spins up. They divide the job among themselves and check each other's changes before anything comes back to you.
Here's a plain-English way to picture it. Hiring one freelancer means you babysit every draft. Hiring a small team means the draft gets reviewed before it lands in your inbox. That's the difference between a normal chat window and `/swarm`.
Seven PDFs went in, a 3D office came out
Google's own example is the part people keep forwarding. Someone handed the swarm seven PDFs of architectural drawings — flat files, the kind an architect emails you — and the agents turned them into an interactive 3D model of an office. Not a rendering. Not a slideshow. A model you can move around inside.
That's a genuinely new kind of output for a general-purpose AI. Until now, "AI plus 3D" usually meant generating a pretty picture of a chair. This means reading a stack of technical documents, understanding how they relate to each other, and assembling something spatial out of them.
What it costs during the intro window
Through the end of November, Gemini 4 Carbon runs at $2 per million input tokens and $10 per million output tokens. A token is a chunk of text — roughly three-quarters of an English word — so a million tokens works out to hundreds of thousands of words going in, or your answer coming back out. Input is the material you feed it; output is what it writes back.
The price is expected to roughly double once that introductory period ends. One honest caveat: plenty of coverage pins the increase to December 1, but Google hasn't confirmed a hard date. Treat the doubling as expected and the calendar as unconfirmed. If you're budgeting, assume November pricing is temporary.
What it means for you
Strip away the jargon and this is the shift: you can hand an AI a pile of messy documents and get back something you can look at from more than one angle. If you've ever squinted at a floor plan on your phone, you already know why that matters.
At home
Renovating? Your architect sends you ten PDFs and you nod along without understanding any of them. A tool like this gives you a rough 3D version you can rotate on a laptop — enough to spot that the new kitchen island blocks the back door before anyone drills a hole. It's a sanity check, not a substitute for the drawings.
At work
Facilities, safety and office-management teams live inside documents nobody reads. Upload the drawings and the spec sheets, ask for a visual summary, and use it to brief people who won't open a PDF. Same logic works for onboarding packs, site plans and equipment manuals.
Running a business
Quoting jobs is where small firms bleed time. If you build, fit out or install, the bottleneck is usually reading the client's plans and picturing what's actually being asked. Feeding a batch of documents to an agent and getting a first-pass visual back can cut hours off a quote — and if you need to tidy those files first, free utilities like the ones at MyKreaTool will compress and merge PDFs without a subscription.
Studying
Architecture, engineering and design students work with thick sets of drawings they can't afford to model by hand. A swarm that reads seven PDFs and returns a rough 3D layout is a first draft you refine — much better than staring at page 40 at 2 a.m. wondering how the sections fit together.
Creative projects
Game designers, set designers and event planners all start from floor plans. Imagine sketching an escape room on paper, scanning it to PDF, and getting back a block-out you can walk through before building anything. The model won't be beautiful, but it will be spatial — and spatial is the hard part.
Earning money
Every new capability creates a small services gap. People who can translate a folder of drawings into a 3D walkthrough, or who can package document-to-model workflows for local firms, have something to sell while everyone else is still reading the announcement. The intro pricing makes a few experiments cheap enough to be worth the risk.
How to try it right now
The cheapest place to start is the Gemini app, because it's the only route with no setup at all. Google hasn't confirmed a free tier for this specific model, so think of the app as your zero-effort check: open it, see whether Gemini 4 Carbon is available on your plan, and ask it a document question you already know the answer to. If it's not there yet, the Gemini 4 Carbon announcement spells out who currently gets access.
1. Check the Gemini app first. No code, no keys. If the model shows up in your model picker, you're in.
2. Need API access? Compare the model listing on OpenRouter to see providers and pricing before you commit, then read Google's introduction post for the setup details.
3. Want the swarm? Head into Antigravity and look for the `/swarm` preview. Remember it's a preview, not a finished product.
4. Prep your files. Put everything in one folder, name files clearly (plan-level-1.pdf, not scan0043.pdf), and delete duplicates. Agents are only as good as the pile you hand them.
5. Write the request in plain language. "These seven PDFs are drawings of a single-floor office. Build me an interactive 3D model showing walls, doors and desks." No prompt engineering degree required.
6. Check the output against the source. Open the PDFs side by side with the model and hunt for missing walls, flipped doors and invented rooms. Every time.
If you want a technical rundown before you spend anything, NeuralTrust's breakdown is a useful second read.
Upsides and what changes
The obvious upside is time. Work that used to mean a human reading dozens of pages and then rebuilding them in CAD software now has a fast, cheap first pass. A rough 3D layout in minutes changes how you quote, pitch and plan.
Less obvious but bigger: outputs get more varied. Chatbots produce text. This produces files you can look at from six angles. Once a model can reliably cross-reference a stack of documents and return something structural, the list of jobs you'd hand it grows quickly — site surveys, fit-out proposals, event layouts, product assembly guides.
And the swarm pattern is quietly important. Multiple agents that check each other's work is a decent answer to the classic AI problem of confident nonsense. It's not a guarantee, but it's a real structural improvement over one model riffing unsupervised.
Limitations
Be clear-eyed about what this is. A 3D model rebuilt from seven PDFs is a sketch, not a construction document — it can miss walls, misread scale and quietly invent a corridor, so nobody should pour concrete based on it. Swarm agents checking each other reduces errors; it doesn't eliminate them, and a group of agents can agree on the same wrong answer. The token math also bites on big jobs: seven detailed PDFs are a lot of input, and if you iterate a dozen times, the bill grows faster than the demo suggests. Access is still gated to developers, enterprises and Google AI Ultra subscribers, so plenty of readers can't just open the app and try it today. And the pricing timeline is fuzzy — the doubling is expected, the December 1 date isn't confirmed by Google, so don't build a budget on a date you read in a newsletter.
Conclusion
The interesting story here isn't a new model number. It's that a Google tool read seven blueprint PDFs and produced an interactive 3D office, and that you can point that same machinery at your own documents while the pricing is still low.
One action for today: open the Gemini app and ask it one question about a PDF you already understand cold. You'll learn in two minutes whether this is hype or a tool you'll actually use next week.



Comments 0