Gemini 4 Carbon Review: Code-Generated Video and 3D Scenes

What videos have people made with Gemini 4 Carbon?#

Community creators have published 181 documented video demonstrations in the community gallery using Gemini 4 Carbon to generate animation code. These projects show how the model converts text prompts into Three.js 3D environments, animated SVGs, and WebGL motion reels. The tables below group prominent community examples by format.

Short films, anime, and narrative animation

Motion graphics and launch showreels

3D architectural scenes and previs

Interactive simulations and environments

You can inspect all 181 published examples in the sandbase-lab/awesome-gemini-4-carbon repository on GitHub, though some entries are cataloged under the Argon codename.

Across these examples, several common characteristics stand out. Every scene is executed as code running directly in the browser, using Three.js WebGL rendering, 2D HTML Canvas, or vector SVG elements. Video runtimes remain mostly under two minutes, which aligns with browser memory performance for complex programmatic animations.

This code-based technique works exceptionally well for procedural 3D environments, motion graphics titles, product exploded views, and stylized visual explainers. However, it cannot generate believable human anatomy or photographic facial performances. When stories demand realistic human actors, creators pair these environmental code passes with dedicated neural video models like Seedance or MiniMax.

Google demonstrated this structural capability directly during launch. Google highlighted an interactive Three.js 3D reconstruction of the Eiffel Tower spanning its 1887 to 1889 construction phases, built from historical photos and engineering plates using Antigravity’s multi-agent /swarm feature. On developer boards, a linux.do creator produced a 15-second motion graphics reel in a single shot by prompting the model to design a dynamic portfolio showpiece.

How do you turn a Gemini 4 Carbon prompt into an MP4 video?#

Converting a Gemini 4 Carbon prompt into an MP4 video requires a four-stage pipeline spanning prompt design, code generation, browser rendering, and screen capture. First, you specify the visual composition, lighting rigs, material properties, and camera moves in a detailed prompt. Gemini 4 Carbon processes the request and writes a single, self-contained HTML file containing all necessary styles, scripts, and runtime logic.

Next, you open that HTML file in a modern browser such as Google Chrome to render the scene at 60 frames per second. As explained in the video tutorial in the community repository, creators capture this real-time execution using screen recording software or automated headless capture scripts. That recording outputs a clean MP4 or WebM video file ready for editing.

Checked on 2026-11-11.

What prompt template produces reliable Three.js video?#

A reliable creative coding prompt uses three mandatory rules: an explicit Three.js r160 CDN link, zero external file dependencies, and an automated camera orbit. Requiring procedural geometry ensures the file runs anywhere without missing texture files.

You can adapt the structured template below to build reliable Three.js video backgrounds and motion graphics:

Create a single self-contained HTML file that renders an animated 3D scene using Three.js (r160 CDN: https://cdnjs.cloudflare.com/ajax/libs/three.js/r160/three.min.js). Do not load external 3D models, textures, or secondary asset files; generate all geometry, shaders, and procedural materials in code.

Visual theme: [Describe the world, e.g., a futuristic floating glass island with bioluminescent plants].
Camera movement: Program an automated, smooth camera orbit that completes a 360-degree rotation every 20 seconds at a fixed elevation angle.
Lighting and atmosphere: Set up a three-point lighting system with a warm key light, a cool rim light, and subtle ambient occlusion.
Animation loop: Animate scene elements continuously with sinusoidal floating motion, rotating parts, and particle movement.
Render requirements: Target 60 FPS on a responsive full-screen canvas, handle window resize events cleanly, and apply tone mapping for vibrant contrast.

This template forces the model to construct the entire scene graph inside one standalone file. Specifying camera timing and lighting coordinates in advance lets the browser animation be recorded directly to MP4 without manual mouse adjustments.

What problems and limits do testers report with Gemini 4 Carbon?#

Early testers report significant operational bottlenecks, including generation times stretching from 16 minutes to over two hours alongside intermittent connection timeouts. When tasked with intricate 3D geometry, the model spends extended periods in deep verification loops. For instance, one linux.do tester reported waiting nearly two hours for a 3D pelican scene while observing generation speeds around 30 tokens per second.

Connection stability during these lengthy reasoning sessions is another frequent complaint among creators. During that same test, the author noted multiple network disconnections before the final HTML file arrived. In another published test, rendering an ink-wash animation took 1 hour and 13 minutes of model compute before finishing.

Developers also express mixed views regarding Carbon’s code aesthetics and user interface design. On TestingCatalog, early testers noted that Carbon is slower but less lazy than Argon, yet cautioned that frontend visual layouts remained basic. Conversely, an open-source Android developer on linux.do praised Carbon’s natural communication and design taste after spending $20 in tokens on UI refactoring.

How do Gemini 4 Carbon and Argon compare on price, speed, and access?#

Gemini 4 Carbon matches Argon’s introductory price of $2 per million input tokens while requiring a $99.99 monthly Ultra plan for chat access. According to 9to5Google, this promotional pricing window closes on November 30, 2026, after which standard rates increase to $4 per million input tokens and $20 per million output tokens. That makes Carbon substantially cheaper per token than competing frontier models during its launch month.

Access tiers create a clear division between the two models. Gemini 4 Argon remains accessible on Google AI Pro plans priced at $19.99 a month, whereas Carbon is restricted to Google AI Ultra subscribers paying $99.99 a month for Ultra 5x or $199.99 for Ultra 20x. On independent benchmark leaderboards, Artificial Analysis ranks Carbon #1 with an index score of 59, one point ahead of Claude Opus 5.5 at 58.

Vendor-reported benchmarks highlight similar gains, with Google reporting top rankings on CursorBench 4.0, WANDR, and AutomationBench at lower cost per task than competing frontier models. On web development leaderboards, Google noted that Carbon gained 89 Elo points to reach fourth place on WebDev Arena. The table below compares current prices and catches:

Checked on 2026-11-11.

For creators deciding between models, Argon remains faster and cheaper for drafting text scripts and simple code. Carbon earns its cost when you need sustained spatial reasoning to assemble intricate 3D scenes without geometric clipping.

Where does Gemini 4 Carbon fit into an AI filmmaking workflow?#

Gemini 4 Carbon provides procedural scene previs and motion design that directors export to video and paste into DramaCanvas alongside dedicated character models. In an AI filmmaking pipeline, you use Carbon to build 3D environments, plan camera angles, and animate opening titles. Once recorded to MP4, these clips serve as structural animatics that establish timing and scene geography.

When your story requires realistic actors, dramatic lighting, and emotional performances, you bring those assets into DramaCanvas. On our infinite canvas, creators organize scene beat sheets and connect specialized video generation models to produce shots with real-looking people. Following our AI short drama workflow guide helps teams maintain character consistency across multi-scene episodes.

DramaCanvas does not offer Gemini models. Generate their HTML animation assets in AI Studio or Antigravity, record the screen, and import the MP4 files into DramaCanvas.

What are the limits of this roundup?#

DramaCanvas ran no benchmark of its own, relying entirely on verified third-party evaluations, vendor technical disclosures, and published community test logs. Synthetic benchmark indices like Artificial Analysis offer general measures of reasoning, but they do not guarantee that every creative coding prompt will compile cleanly in a browser. Real-world performance depends heavily on local GPU acceleration, browser WebGL support, and prompt specificity.

Community demonstrations on social media represent curated successes that often conceal repeated prompt adjustments and failed attempts. While recorded clips show impressive visual fidelity, they do not reflect guaranteed one-shot completion rates for inexperienced users. Furthermore, token economics will change once Google’s introductory promotional rates conclude at the end of November 2026.