39+ Gemini 4 Carbon 3D Scenes, Blender and Native Code Prompts
39+ prompts. 223 things people built, tested and compared with Gemini 4 Carbon in its first weeks: 3D scenes, games, simulations, films made from code, real bug fixes and research, and dozens of same-prompt tests against Claude Fable 5.2 and Gemini 4 Argon. Each is credited and linked to the original post, with the prompt where the creator published one.
Gemini 4 Carbon (model id gemini-4-carbon) is Google's deeper-reasoning Gemini 4 model. Google built it on Gemini 4 Argon and describes the difference as more thinking tokens per turn; it opened to vetted cybersecurity defenders on 28 October 2026 and to every paid Gemini API developer on 10 November. It reads text, images, video, audio and PDFs with a 1,048,576-token context and writes text, so the scenes, games and films here are code it wrote (Three.js, WebGL, Blender Python, CUDA C++, Web Audio), mostly run by Google's Antigravity agent. Of the 223 cases, 18 include the prompt as published, 33 only the creator's request and 172 no prompt; we never fill one in. In same-prompt tests creators usually found Claude Fable 5.2's results more polished and Carbon's cost a small fraction of Fable's. On apimodels.app Carbon costs $1.10 per 1M input tokens and $5.50 per 1M output through 30 November 2026 and $2.20 / $11 from 1 December, against Google's $2 / $10 and $4 / $20.
Mechanical flower in Three.js with an unprompted bee
Prompt not published
No prompt published. A mechanical flower built entirely in Three.js in about 1 hour 43 minutes. The author only asked it to be creative without constraints, and it added a small bee that was not in the prompt; the author saw many improvements over Argon.
Clip:
Video, 44 s. About 59K views and 1,390 likes on X as of 12 November 2026.
Numbers the author reported:
~1hr 43mins
Note:
Posted 21 Oct, before the 10 Nov API launch. A week before Carbon was announced on 28 Oct; the post does not say how the author had access.
No prompt published. With early access, the author gave Carbon a video clip and it reproduced the scene in Three.js. The author calls Carbon a very good model that improves on Argon and thinks differently from Claude and GPT.
Clip:
Video, 14 s. About 61K views and 977 likes on X as of 12 November 2026.
Note:
The thumbnail shows the source, a train-cab clip credited to Dr. Akiyoshi Kitaoka, next to the Carbon recreation.
No prompt published. The first of 13 coding briefs (3D, websites and apps) the author ran with early access in Antigravity. One control drives the orrery's gears and folding petals, even in reverse, and one follow-up improved the framing.
Clip:
Video, 15 s. About 53K views and 717 likes on X as of 12 November 2026.
Tool:
Antigravity
Numbers the author reported:
13 coding briefs
Note:
First post of a thread; the other briefs are in the replies. @xbquan reposted another clip from it: a web page running 1.04 million particles from one prompt.
Open valley scene in custom C++ with no reused assets
Prompt not published
No prompt published. A scene made by Carbon on High reasoning effort in custom C++ with no reused assets, over two prompts; the author says the first result was already good enough and the second pushed it further.
Clip:
Video, 70 s. About 9,430 views and 337 likes on X as of 12 November 2026.
Numbers the author reported:
two rounds/user prompts
Note:
Subject (a green valley with a stone bridge) is from the thumbnail. @laopai1990 and @Sin_Yiyao reposted the same video; the former likens it to Zelda: Breath of the Wild.
No prompt published. With early access, the author built a 3D basketball court with no premade 3D assets. The author found Carbon much better than Argon at anything involving taste and visuals: 3D, websites and motion graphics.
Clip:
Video, 30 s. About 8,904 views and 308 likes on X as of 12 November 2026.
Jeweled sea creature sculpted and animated in Blender
Prompt not published
No prompt published. A 3D test that Argon failed weeks earlier, rerun in Antigravity on High. Carbon took 81 min and 186k output tokens (about $4.5 in API terms) versus Argon's 14 min, 11k tokens and about $1.8; the piece was made with the Higgsfield plugin in Blender.
Clip:
Video, 30 s. About 13K views and 295 likes on X as of 12 November 2026.
Tool:
Antigravity
Numbers the author reported:
Gemini 4 Carbon: 81 min, 186k output tokens, about $4.5 in API terms · Gemini 4 Argon: 14 min, 11k output tokens, about $1.8 in API terms
Compared with:
Gemini 4 Argon. Carbon delivered where Argon had left the author highly disappointed.
Note:
Subject (an ornate, jeweled seahorse-like creature) is from the thumbnail, which also reads "same prompt, rigged and animated in Blender".
No prompt published. A grass meadow built from scratch in CUDA and C++ with no game engine or 3D library: 52k blades plus wildflowers. The wind sound is also synthesized in C++, so the audio follows the same gusts as the grass.
Clip:
Video, 5 s. About 109K views and 200 likes on X as of 12 November 2026.
Island castle and dragon: first prompt vs three follow-ups
Prompt not published
No prompt published. The same island, castle and dragon after Carbon's first prompt and after three more follow-ups: in the first pass the fire is nine orange cubes, after the follow-ups it looks almost real. The author's takeaway is that the gap comes from the follow-ups, not the model.
Clip:
Video, 13 s. About 25K views and 178 likes on X as of 12 November 2026.
Aquarium with angelfish, snail and shrimp from a multi-tool pipeline
Prompt not published
No prompt published. Carbon built an entire aquarium, including the angelfish, snail and shrimp, using Blender, Three.js, build123d and a local image-to-3D pipeline. Total time was about 6 hours.
Clip:
Video, 64 s. About 4,332 views and 87 likes on X as of 12 November 2026.
Numbers the author reported:
approximately 6 hours
Note:
Part of the author's thread of 7 builds made with early access.
Koi courtyard with water reflections and swimming fish
Prompt not published
No prompt published. The author's Koi Courtyard test: the environment, water reflections, lighting and swimming koi all impressed the author. The only downside noted is that Carbon is a bit slower than Argon.
Clip:
Video, 42 s. About 3,777 views and 81 likes on X as of 12 November 2026.
Note:
@AnikaCodes reposted it, describing ripples on click, a Feed button that brings the koi over, and a camera that dives underwater.
No prompt published. Carbon recreated a Coca-Cola bottle with math alone: no 3D models, no external assets, everything in a single shader, and the cap has all 21 teeth.
Clip:
Video, 30 s. About 3,927 views and 74 likes on X as of 12 November 2026.
Minecraft-style test in Three.js using Antigravity voice mode
Prompt not published
No prompt published. First results from the author's Minecraft test, built with Three.js in Antigravity and driven partly by voice mode. The author says Carbon came close to Opus 5.5 on this test and now wants to try harder agent and coding tasks.
Clip:
Video, 65 s. About 2,376 views and 73 likes on X as of 12 November 2026.
No prompt published. Made while Carbon was shadow-routed under the name gemini-4-argon: 11,286,682 voxels at 60fps in one standalone HTML file, with moving viaduct trains, sky-harbor airships and three live lighting modes.
Clip:
Video, 45 s. About 2,242 views and 66 likes on X as of 12 November 2026.
Numbers the author reported:
11,286,682 voxels at 60fps
Note:
Pre-release: made while Carbon was served under the gemini-4-argon model name.
Destroyed Coventry Cathedral rebuilt from floorplans and photos
Prompt not published
No prompt published. Carbon rebuilt the destroyed Coventry Cathedral from scratch by reading historic floorplan drawings and photographs, then generated a cinematic walkthrough video.
Clip:
Video, 100 s. About 6,870 views and 51 likes on X as of 12 November 2026.
Procedural mechanical falcon with a working watch movement
Prompt not published
No prompt published. Built on early access: a fully procedural Three.js mechanical falcon with a live 28,800 VPH watch movement in its chest, a 4-bar wing rig and a 240 km/h dive. No 3D assets or external textures.
Clip:
Video, 35 s. About 1,386 views and 45 likes on X as of 12 November 2026.
Voxel floating kingdom with over three million voxels
Summary
Only the creator's request is public, not a full prompt: "Voxel Floating Kingdom"
What it is:
A voxel floating kingdom with over 3 million voxels, made on high reasoning in Cursor in about 1 hour 15 minutes. The author puts the API cost at about $0.95 for an estimated 11.7M tokens.
Clip:
Video, 54 s. About 1,732 views and 35 likes on X as of 12 November 2026.
Tool:
Cursor
Numbers the author reported:
~ 1h 15m · ~ $0.95 in API costs · over 3 million voxels · est. 11.7M tokens
Note:
Cost and token figures are the author's estimates. A follow-up post (2120479775764459951) compares the same run with Opus 5.5.
Eiffel Tower model built from old drawings and photos
Prompt not published
No prompt published. On launch day the author shared an Eiffel Tower model built with Carbon from old drawings and photos, saying the team found that Carbon's visual reasoning makes it strong at turning 2D into 3D.
Clip:
Video, 59 s. About 1,497 views and 33 likes on X as of 12 November 2026.
Note:
The author writes "we discovered", which suggests a Google team member. @choblin29 posted a quoted description of the same demo next to Sundar Pichai's launch post: Antigravity /swarm agents turned historical photos and engineering plates into an interactive Three.js exhibition of the 1887-1889 construction from a single prompt.
No prompt published. A space video the author made with Carbon using only code and Blender. The author argues Carbon is no worse than Fable 5.2 and is fast, capable and uses fewer tokens.
Clip:
Video, 35 s. About 2,475 views and 30 likes on X as of 12 November 2026.
Note:
The thumbnail caption reads "Blender MCP · procedural modeling · zero external assets".
Supercar chasing a steam train, scripted in Blender
Prompt not published
No prompt published. Part of the author's tests of Carbon on Blender scripts: from one prompt, a supercar chasing a steam train with wet-road reflections and camera cuts, which came out better than the author expected. It took about 1h54m and $11.80.
Clip:
Video, 10 s. About 2,122 views and 13 likes on X as of 12 November 2026.
Numbers the author reported:
~1h54m ($11.80)
Note:
Posted 8 Nov, before the 10 Nov API launch. Model listed as carbon-h.
Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stomps around manhattan and queens and Brooklyn, use openstreetmap mapping data
What it is:
Simon Willison asked Carbon in Antigravity for a Three.js animation of Yonkers standing up as a mech and stomping around Manhattan, Queens and Brooklyn using OpenStreetMap data. Willison also ran the pelican SVG at low, medium and high thinking, and had to add "build that into a standalone HTML file" to get a page with no external dependencies.
Clip:
About 410 views and 12 likes on X as of 12 November 2026.
Tool:
Antigravity
Note:
Made by Simon Willison (@simonw, original post https://x.com/simonw/status/2120258951594037732); shared with a Chinese summary by @ZSS_HK, whose post is the url here. Follow-up line used: "build that into a standalone HTML file".
Icelandic black-sand beach in Three.js over five rounds
Prompt not published
No prompt published. An Icelandic black-sand beach made from scratch in Three.js with no external assets; every basalt column, wave, sky and wet-sand reflection is generated in code. After 5 rounds the sky gradient, sea-stack reflections and wave bands worked, but the basalt close-ups still looked like plastic.
Clip:
Video, 54 s. About 757 views and 10 likes on X as of 12 November 2026.
Endless marble-run tower where every step plays a note
Summary
Only the creator's request is public, not a full prompt: "a marble run"
What it is:
Asked for a marble run with /swarm on, Carbon built an endless tower in one Three.js file: marbles roll through glass tubes, spin a paddlewheel and bounce down neon xylophone stairs that each play a note, then a lift carries them back up.
Clip:
Video, 15 s. About 1,072 views and 9 likes on X as of 12 November 2026.
Note:
The post mentions /swarm but does not name the tool.
No prompt published. The author modeled a barbecue stall with Carbon in Antigravity and felt the details were somewhat better than an earlier attempt with Argon.
Clip:
Video, 47 s. About 1,860 views and 7 likes on X as of 12 November 2026.
Rifle assembly and ballistics scene in Blender with no downloads
Prompt not published
No prompt published. Using only Python, Blender and Carbon with zero downloaded 3D assets: a part-by-part AK-47 assembly, live fire-mode switching, supersonic bullet-time and a ballistic watermelon destruction with custom foley audio.
Clip:
Video, 18 s. About 520 views and 7 likes on X as of 12 November 2026.
Note:
An earlier post by the same author (2120443587100049782) shows the Blender viewport mid-build.
Build a 5,000-drone light show over a night waterfront. One HTML file. No external assets.
What it is:
From one prompt in /swarm mode: 5,000 glowing drones in a single InstancedMesh at 60 FPS, with neon bloom, a skyline and reflective water. The drones swirl between four formations (spinning Earth, DNA helix, rainbow torus knot, flapping phoenix), with live FPS, formation buttons and sliders.
Clip:
Video, 30 s. About 906 views and 6 likes on X as of 12 November 2026.
Tool:
Antigravity
Note:
The quoted earlier post by the same author shows a playable Three.js Rubik's Cube that stays aligned after 100+ turns.
No prompt published. A swarm of Carbon agents built a hypercar in Three.js on its own in about an hour. The author calls it a step up from previous models, though the car looks a bit damaged.
Clip:
Video, 85 s. About 781 views and 5 likes on X as of 12 November 2026.
No prompt published. A sailing-boat test finished in 9 minutes in the Gemini mobile app. Materials, splashes and lighting are solid, and the model added many adjustable parameters on its own.
Clip:
Video, 97 s. About 545 views and 5 likes on X as of 12 November 2026.
Pure CUDA + C++20 ray tracer: raindrops on a lotus pond
Prompt not published
No prompt published. Carbon wrote a standalone macro optical ray tracer, shallow-water and droplet GPU physics and an audio synthesizer that render a 20 s 1080p60 clip with 48 kHz audio, with no engines, 3D libraries or external assets.
Where:
GitHub, 2026-11-11
Tool:
Google Antigravity with Gemini 4 Carbon
Numbers the author reported:
3,899 lines of CUDA/C++20 · renderer 1,431 lines · physics 737 lines · 65,536 particles · 1024x1024 pond grid · 20 s 1080p60 clip with 48 kHz audio
Note:
The author also posts on X, so this may overlap with X-sourced cases.
181-part Song Dynasty mortise-and-tenon pavilion with exploded view
Prompt not published
No prompt published. In one HTML file, Carbon built a millimetre-scaled pavilion from the 1103 AD Yingzao Fashi, with spring disassembly, a rain-night scene, a Kuwahara ink-painting shader and Karplus-Strong knock sounds. A live demo is online.
Where:
GitHub, 2026-11-11
Tool:
Google Antigravity with Gemini 4 Carbon
Numbers the author reported:
4,525 lines · 181 meshes
Note:
The author also posts on X, so this may overlap with X-sourced cases.
Create an SoC linux based motor controller using tscircuit
What it is:
In a YouTube short, both models were asked to design an SoC Linux-based motor controller in tscircuit. The author judged Fable slightly better but 20x more expensive ($200 vs $12).
Where:
YouTube video, 2026-11-11
Tool:
tscircuit
Numbers the author reported:
$200 vs $12 (Fable vs Carbon) · Fable 20x more expensive
Compared with:
Fable 5.2. Fable is slightly better but 20x more expensive.
Note:
The author adds that Carbon is better than Astra at PCB design, which this video does not show.
Reference-image 3D modelling and a Rube Goldberg machine vs Opus 5.5
Prompt not published
No prompt published. Carbon, through Factory AI's API (two accounts, each task tried twice), took on reference-image 3D modelling and a Rube Goldberg machine against Opus 5.5 in Claude Code at extra high, with DeepSeek 4.1 on the DSH platform as a floor. The video itself was made by Carbon.
Where:
Bilibili video, 2026-11-11
Tool:
Factory AI's API
Numbers the author reported:
two accounts, each task tried twice
Compared with:
Opus 5.5, DeepSeek 4.1. Carbon did not beat Opus 5.5 on these two tasks but was clearly stronger than DeepSeek 4.1.
No prompt published. Carbon at medium effort drove Blender for 6 hours to build a 3D model of Paimon. The author says the facial expression is not good but stopped the run because of token cost.
1Write a spec, not a wish. Creators who gave Carbon one vague line mostly got average results; one who ran 15 projects summed it up as: don't make wishes at it. The striking results came from briefs that named the stack, the deliverable and what done looks like, up to a spec of about 10,000 characters for a three-minute violin concert film.
2Show it a reference. Carbon reads video. One creator gave it the gameplay video of a shooter Claude Fable 5.2 had built and asked for the same game or better; another gave it Anthropic's own Fable 5.2 demo video to rebuild as a web page.
3Make it look at its own output. The shooter rebuild ran a headless-Chrome screenshot loop that caught and fixed three bugs on its own, in 48 minutes for $3.82. Another creator had Carbon review its screenshots like a VFX supervisor four times, about two hours and $7.40, turning a block of voxels into a convincing dinosaur.
4Name the format. "One HTML file. No external assets." is half of the 5,000-drone light show prompt. Simon Willison had to add "build that into a standalone HTML file" to get a page he could open on its own.
5Split big builds across agents. The largest builds used Antigravity's multi-agent /swarm mode: nine agents and 22,910 lines for a diorama builder in 1 hour 41 minutes, and five waves of agent critique for an open-world game.
6Follow up. One creator posted the same island and castle after the first prompt and after three follow-ups; the fire went from nine orange cubes to something close to real. The gap, in the author's words, came from the follow-ups.
7Check what it says it finished. In one six-app test every one of Carbon's final summaries claimed at least one feature that turned out broken or overstated. Ask it to click through the result in a browser and report what it saw.
Prompt formula
Carbon build brief = the deliverable and format (one HTML file, an MP4, a repo) + the stack and constraints + what done looks like + a reference (an image or video it should match) + a self-check (open it in a browser, screenshot, fix) + a stop condition or time and token cap
Example (our template, not a creator's prompt): Build a browser game: a night-time arcade racer on a neon highway. Three.js from a CDN, music and sound in Web Audio, everything else in one index.html. Match the mood of the attached video. Done means: it loads with no console errors, steering and crashes work, and a lap takes about 90 seconds. Open it in a headless browser, take screenshots at the start and mid-lap, fix what looks wrong, then tell me what you checked. Do not ask questions.
Frequently asked questions
Can Gemini 4 Carbon make 3D scenes, games or videos by itself?
Not directly. Carbon takes text, images, video, audio and PDFs and returns text. Every scene, game and film here was rendered from code it wrote: Three.js and WebGL pages, Blender Python, CUDA or C++ programs, Web Audio for sound. In a chat window you get the code or an HTML file; an agent such as Antigravity, Cursor or GitHub Copilot can run, render and test it.
Which tool did people use with Carbon?
Mostly Google's Antigravity, often with its /swarm multi-agent mode. Others used Cursor, GitHub Copilot, AI Studio, Vertex AI or the API from their own harness. Where the creator said, the card names the tool.
Is Carbon better than Claude Fable 5.2 for this kind of work?
Usually not on polish, often yes on cost. In most same-prompt tests here Fable 5.2 produced the better-looking, more complete result, while Carbon cost a fraction as much and sometimes won outright: $7.54 against $41.71 for three jobs in one test, where Carbon won two of the three.
Is Carbon better than Gemini 4 Argon?
Mostly yes, at the cost of time and tokens. The two cost the same per token; in these tests Carbon usually produced the stronger result and took several times longer, for example 81 minutes against 14 in one Blender test that Argon had failed. Google still recommends Argon for most everyday coding.
How much did these builds cost?
As the creators reported: $0.45 for a one-shot 3D platformer of 2,599 lines, $3.82 for the shooter rebuild, about $7.40 for four rounds of self-review and $27.68 for three long builds. Cost depends on the thinking level and how long the agent runs; thinking tokens are billed as output.
How do I run Gemini 4 Carbon?
On apimodels.app, call model gemini-4-carbon on the OpenAI-compatible /v1/chat/completions endpoint or the native Gemini endpoint with one key, from anywhere including mainland China. Raise the output limit to at least 32,768 tokens, stream long answers and allow a 15-minute timeout, because Carbon thinks for a while before it writes.
Can I reuse these prompts and videos?
The works belong to their creators. Clips from X play from X's own servers, videos on YouTube and Bilibili play in their own players, and every card links to the original; we do not host or re-upload any of them. Published prompts are quoted with credit, in full when short and excerpted when long. If you made one of these and want the card removed or the credit corrected, email us.
Run these prompts on apimodels.app
Gemini 4 Carbon is live here alongside 160+ other image, video, audio and language models — one API, one key, pay as you go, with $0.10 of trial credit on sign-up.
The whole collection is on GitHub as Markdown and JSON: all 223 cases with credits and links to the original posts and pages, the tools and numbers each creator reported, and a summary of the same-prompt tests. The videos and images stay with their creators and are linked, not re-hosted. Prompts up to 400 characters are quoted in full; longer ones are excerpted, and unpublished ones are left out. Our own write-ups are licensed CC BY 4.0. Open an issue to add a case or correct a credit.
