Skip to content
Theme
OnDevice Tools

AIAI Models & Research

Gemini 4 Carbon: five developer projects show what it can do

Google's Gemini blog describes a wind-blown meadow, a million-particle simulation and a stained-glass stacking game among the first public builds. The model can be reached through the Gemini API and Google's developer tools.

Illustration of a Gemini 4 Carbon screen showing a simulated meadow and particle cloud

Key takeaways

  • Google calls Carbon its strongest model so far for complex technical work.
  • One build simulates 1,048,576 glowing particles in real time using WebGPU.
  • Another turns scattered public records into 14 interactive tools with 3D graphics.
  • Carbon is offered through the Gemini API, Google AI Studio, Android Studio and business integrations.
In this article
  1. What Google announced
  2. A meadow, a tidal current and stained glass
  3. Public records and a Parisian laboratory
  4. How to try Gemini 4 Carbon
  5. What the showcase does not prove

What Google announced

Illustration: Google logo, AI, Gemini logo, code and the labels Cuda, Gemini 4 Carbon, Gemini API

Google has published five projects built with Gemini 4 Carbon, the company's newest model for demanding technical work. The showcase went up on the Gemini blog on 17 November 2026, about a week after Google said it had opened the model to a wider group of developers.

Google describes Carbon as its most capable model for complex technical tasks, and says it is built to reason at length and to keep checking its own results. The company says the community has been running long verification loops, in which the system examines a result again and again before accepting it. Google points to physics simulations, interactive data tools and detailed virtual environments as the areas where builders are trying the model. No benchmark scores, prices or regional availability details appear in the post.

The model can be reached through the Gemini API, Google AI Studio, Android Studio and business integrations, according to Google. The Gemini API is the connection developers use to call a model from their own software.

A meadow, a tidal current and stained glass

The first project is a meadow. Glitter Gal, who posts as GlitterPixely on X, wrote a field of 52,000 flowers and blades of grass from scratch in CUDA and C++, with help from Carbon. CUDA is a programming framework that runs heavy calculations on graphics hardware. The wind in the scene was generated to match each gust.

Chubby, known as kimmonismus on X, wanted to see more than a million glowing particles drift and turn like tidal water. Using Carbon with WebGPU, a browser technology for 3D graphics, the build animates 1,048,576 particles in real time.

Andrew V, who posts as AI_Andrew on X, asked how a block-stacking game would change if every block were fragile stained glass. Carbon modelled the way such a tower catches the light, rings when it is tapped and breaks into shards when it falls. Each of the three builds was shared publicly on X by the developer who made it.

Public records and a Parisian laboratory

Giyu_codes went through untidy public records looking for industries that had been left without good software. Working with Carbon, the developer turned scattered datasets into 14 interactive tools with live 3D graphics, and collected all of them in a thread on X.

The Bugged Dev, posting as thebuggeddev on X, began with a single reference image. Carbon helped rebuild it as a Parisian optics laboratory from the 19th century, fitted with 92 instruments made of brass and glass that can be tested and adjusted in the scene. An optics laboratory is a workspace for experiments with light, lenses and mirrors.

How to try Gemini 4 Carbon

Google says developers can start building with Carbon through the Gemini API, Google AI Studio, Android Studio or a business integration of their choice. Google AI Studio is a browser workspace for writing prompts and trying models. Android Studio is Google's environment for building Android apps.

The blog post does not list prices, usage limits or the countries where Carbon is offered. Anyone planning to build on it should check Google's official developer documentation for current terms.

What the showcase does not prove

All five builds were chosen by Google and described by the developers who made them. They show what the model can do in the hands of experienced programmers, but they are not independent tests, and Google published no figures for speed, accuracy or cost.

Teams considering Carbon for production work will need their own evaluations on their own data. The showcase is best read as a sample of the work Google hopes the model will be used for.

Sources

  1. See what 5 builders are making with Gemini 4 CarbonGoogle Gemini

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.