APIMODELSAPIMODELS

Gemini 4 Carbon Agents, Research and Writing 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.

← All 223 Gemini 4 Carbon prompts

Last updated: · Gemini 4 Carbon API — pricing & playground · API docs

TL;DR

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.

Source: Google · Gemini 4 Carbon (the model vendor’s official site)

Hover to preview · click for sound

Agent teams find new exact solutions for fusion equations

Prompt not published

No prompt published. The author's first paper at Google DeepMind: dozens of Carbon agents worked on the differential equations of nuclear fusion devices and found new exact solutions, including one known to exist since 1970 that nobody had found.

Clip:
Video, 9 s. About 248K views and 4,464 likes on X as of 12 November 2026.
Note:
The author works at Google DeepMind, per the post.
video by @lucasjvdsSource
Hover to preview · click for sound

Carbon splitting a scene into tasks for a local model

Prompt not published

No prompt published. In Atomic Agent Fusion, Carbon split the prompt into tasks and sent them to a local Qwen 3.8 27B. The minion scene took 5 minutes and $2.39, versus 16 minutes and $7.71 with Carbon alone.

Clip:
Video, 30 s. About 13K views and 140 likes on X as of 12 November 2026.
Numbers the author reported:
5 min and $2.39 · 16 min and $7.71
Compared with:
Gemini 4 Carbon alone. Orchestrating a local model was about 3x faster at about a third of the cost.
Note:
Promotional post by the Atomic Agent team, run in their own orchestration product.
Hover to preview · click for sound

Racing language models in a sailing simulator

Prompt not published

No prompt published. The author raced the new Carbon in a sailing simulator. Carbon did best across multiple runs, while Argon had the single fastest run in the race shown; the author wonders why these models sail better than Fable and Astra.

Clip:
Video, 38 s. About 14K views and 121 likes on X as of 12 November 2026.
Compared with:
Gemini 4 Argon, Claude Fable (version not stated), GPT-6 Astra. Carbon did best across runs; Argon had the single fastest run.
video by @shekitupSource
Hover to preview · click for sound

Fourteen vertical data apps from public datasets

Summary

Only the creator's request is public, not a full prompt: "find high value underserved industries with publicly available data so we can build full vertical platforms"

What it is:
Using four of the author's existing public-data apps as architectural guidance, Carbon was asked to find high-value underserved industries with public data and build full vertical platforms. All 14 apps were built into hosted applications; the author was most impressed by its multi-source data reconciliation.
Clip:
Video, 85 s. About 31K views and 94 likes on X as of 12 November 2026.
Numbers the author reported:
all 14 apps · 4 of my existing apps
Note:
Early access. The author says the build prompts are in the thread replies.
video by @giyu_codesSource
Hover to preview · click for sound

Three cube moves performed with robot hands

Prompt not published

No prompt published. A robot-hand task of three cube moves (playback at 2x): Carbon went 3/3 in 64 minutes for $3.29 and Claude Fable 5.2 went 3/3 in 28 minutes for $6.05. The video ranks 18 models at the end.

Clip:
Video, 38 s. About 499 views and 7 likes on X as of 12 November 2026.
Numbers the author reported:
Gemini 4 Carbon: 3/3, 64 min, $3.29 · Claude Fable 5.2: 3/3, 28 min, $6.05 · All 18 models ranked
Compared with:
Claude Fable 5.2. The verdict is only in the video.
Hover to preview · click for sound

Five-finger hand picking up an apple by its stem

Prompt not published

No prompt published. One attempt each at lifting an apple by the stem with a five-finger hand: Carbon lifted it at 6.9 s (33 min, $2.30) and Claude Fable 5.2 at 8.7 s (24 min, $3.76). A 20-model leaderboard closes the video.

Clip:
Video, 25 s. About 746 views and 6 likes on X as of 12 November 2026.
Numbers the author reported:
Gemini 4 Carbon: lifted at 6.9 s, 33 min, $2.30 · Claude Fable 5.2: lifted at 8.7 s, 24 min, $3.76 · 20-model leaderboard
Compared with:
Claude Fable 5.2. The verdict is only in the video.
Hover to preview · click for sound

Moving an apple over a wall onto a plate

Prompt not published

No prompt published. Five starts each at moving an apple over a wall onto a plate: Carbon went 5/5 in 17.5 minutes for $0.75 and Claude Fable 5.2 went 5/5 in 17.3 minutes for $2.14. The video ranks 14 models.

Clip:
Video, 33 s. About 499 views and 6 likes on X as of 12 November 2026.
Numbers the author reported:
Gemini 4 Carbon: 5/5, 17.5 min, $0.75 · Claude Fable 5.2: 5/5, 17.3 min, $2.14 · All 14 models ranked
Compared with:
Claude Fable 5.2. The verdict is only in the video.
Hover to preview · click for sound

Real robot told to fold a towel, without success

Fold the towel

What it is:
The team tested Carbon on a real robot with the simple instruction "Fold the towel" and waited a long time without success; they plan to try Gemini Robotics 2.5 next.
Clip:
Video, 32 s. About 344 views and 5 likes on X as of 12 November 2026.
Note:
A failure case. The thumbnail shows a physical robot arm over a towel on a table.
by @sanatemSource
哔哩哔哩

Five web-novel openings across five genres

Prompt not published

No prompt published. Carbon wrote five web-novel openings (xuanhuan, apocalypse hoarding, rural urban, sweet-romance quick-transmigration, horror) in 马良写作, and the creators critiqued hooks, character detail and "AI tone" line by line without editing the text.

Where:
Bilibili video, 2026-11-11
Tool:
马良写作
Note:
The author is promoting the 马良写作 writing tool.
Also:
https://www.bilibili.com/video/BV1KfWm6rE95
video by rbbbcSource
One hard Sudoku, five tries each: Carbon vs ArgonYouTube

One hard Sudoku, five tries each: Carbon vs Argon

Prompt not published

No prompt published. Through the Gemini API (Interactions) at thinking_level high, Carbon and Argon each tried one hard Sudoku five times, with answers checked by code. Both got 5/5 right; Carbon was a little quicker, with tokens and cost about the same.

Where:
YouTube video, 2026-11-11
Tool:
Gemini API (Interactions), thinking_level high
Numbers the author reported:
both 5/5 correct · minutes per try: Carbon 6.3-9.6 (median 6.6) vs Argon 6.4-13.5 (median 8.2) · thinking tokens avg 46,857 vs 47,960 · cost avg $0.56 vs $0.57 per try · ten tries about $5.60
Compared with:
Argon. Carbon was a little quicker; tokens and cost were about the same.
Note:
Tested 2026-11-10, published 2026-11-11. The prompt is the PROMPT constant in the published script on GitHub.
Also:
https://github.com/gnana70/gemini-4-carbon-sudoku-test

Reviewing an unfinished paper draft against other top models

Prompt not published

No prompt published. The author fed the same unfinished paper draft to several top models through the API, and Carbon raised points all the others missed. In their words: "It's an anecdote, sample of one, but it got my attention."

Where:
Hacker News, 2026-11-10
Tool:
API
Compared with:
Astra, Fable, Opus, Argon. Carbon raised points all the others missed.
Note:
Anecdotal: a single draft.
Watch on the original post
video by 5555watchSource

How to write great Gemini 4 Carbon prompts

  1. 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.
  2. 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.
  3. 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.
  4. 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.
  5. 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.
  6. 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.
  7. 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.

This library is open source on GitHub

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.

View on GitHub

Other prompt libraries