Make games with Gemini, three ways

By GamesByAI · Published · Updated · 15 sources

On this page

Gemini is Google’s family of AI models, so when you make a game with it, you’re really using one of three Google tools: Canvas in the Gemini app, Build mode in Google AI Studio, or Google Antigravity.

Each of them has a free start, and which one fits mostly depends on where you want the game’s files to end up. The tool facts below are as of September 2026, and each one links to its source.

Pick a route by where the game should live

If you just want a first playable page, start in Canvas. Build mode suits a game you’ll keep editing and publish from Google, and Antigravity is the one to use when you want your own files and engine.

Route What it is Cost to start Good for
Canvas in the Gemini app A workspace that turns a prompt into a working, shareable app or game Included in the free Gemini plan A first playable page and a share link
Build mode in Google AI Studio A prompt-to-app builder with live preview, code editing and GitHub sync AI Studio usage is free of charge A game you refine by chatting, then deploy or download
Google Antigravity An agentic development platform: an IDE with an agent manager, a CLI and an SDK Free for individual developers, with weekly limits Your own project folder and engine

The table draws on Google’s pages for Canvas, Gemini plans, Build mode, Gemini API pricing and Antigravity.

If you’re wondering where Gemini CLI went, it turns out Google’s earlier terminal agent is being replaced by Antigravity CLI (Google Developers Blog). The FAQ at the bottom has the dates.

Start with a playable page in Canvas

Canvas is the quickest way to see something move. You describe a game, Gemini writes the code, and the game runs in a panel right beside the chat. Google says it generates “the code that brings your idea to life as a working, shareable app or game” (Canvas).

Gemini Apps Help walks you through it in four steps.

  1. Go to gemini.google.com, click Add Files below the text box and choose Canvas.

  2. Type your prompt and send it, and the game shows up in the Canvas panel.

  3. When something breaks, click Show console in the Canvas panel to see errors and logs from the preview. If you’d like to see what changed in the code, click Code and then Show recent changes.

  4. To share it, click Share & export, then Share, and copy the link.

Read Google’s warning before you get to step 4, because it’s a little more serious than it sounds. Anyone with a public link can view and edit data associated with the app, and shared apps can save data across sessions and share it between users.

Keep refining it in AI Studio

Build mode is for a game you want to keep working on. It turns a prompt into a full-stack web app, or even a native Android app, with a live preview next to the code, and Google’s docs say the Antigravity Agent powers it (Build mode).

  • Web apps get a React front end and a Node.js runtime for API calls, databases and npm packages. That’s handy for games, because the Gemini API key stays a server-side secret. A game that calls Gemini while you play, say for a talking character or a level generator, doesn’t give the key away.

  • Android apps come out as Kotlin and Jetpack Compose projects. You preview them in a browser-based emulator, install them on a phone and publish them to the Play Store.

From there you keep going by chatting, by annotating the interface or by editing code in the Code tab. If you want the code in a repository of your own, two-way GitHub sync is built in.

When it’s ready, you can deploy to Google Cloud Run, where pricing applies, or download a ZIP and host it yourself. And if you’d rather carry on locally, since Google I/O 2026 one click moves the whole project into Antigravity (I/O 2026).

Keep in mind that Gemini calls from a shared app count toward your usage limits, and paid models may cost money.

Let Antigravity work on your own files

Antigravity is the route for when the game lives in your own folder, with your own engine. Google calls it “our agentic development platform”, and you can use it as an IDE with an agent manager and artifacts, as a terminal-first CLI or through an SDK. There’s also Antigravity 2.0, for running several local agents in parallel (Antigravity).

The launch post describes agents that write the code, use the terminal to launch the application and then use the browser to test it. Along the way they leave artifacts such as task lists, implementation plans, screenshots and browser recordings. It’s a pretty sensible idea, since you can check what they did without reading tool logs.

Google’s own demo is a 3D wizard game set in a castle, with puzzles, and its textures were generated with Nano Banana. According to Google, “Gemini 3.8 Flash built this game with a simple prompt using a looping instruction in Google Antigravity” (Gemini 3.8 Flash post). The prompt section below borrows that looping idea.

The free plan for individuals has basic weekly rate limits, and Google AI Pro and Ultra add more generous limits and an AI credit pool. The agent doesn’t even have to run Gemini, because it can also run Claude Sonnet and Opus 4.6 or gpt-oss-120b, so an Antigravity game isn’t always a Gemini game (pricing).

Give Gemini a small game with a finish line

Start with one small game that Gemini can finish in a single answer. You can paste this into Canvas as it is. In Build mode you can drop the first sentence, since Build mode makes a React project anyway.

Build a browser game in one HTML file. The player is a robot vacuum in a top-down room and steers with drag or arrow keys. Collect dust for 60 seconds while dodging a rolling ball, then show a score screen. It must run on a phone and use relative paths only.

After that, change one thing per message, like “the ball speeds up every 15 seconds” or “add a second room behind a door that opens at 20 dust”, so it’s a lot easier to tell which change broke something.

In Antigravity, end the prompt with a check loop in the spirit of Google’s demo, so the agent looks at its own work before it hands it back.

After each change, launch the game, play it in the browser, take a screenshot and fix what looks wrong before you report back.

Try it on a phone before you share it

Before you send the link around, play through a round on a phone with touch controls and check that you can restart it. Once that works, a Canvas share link is the quickest way to show a friend.

If you’d rather host it yourself, itch.io takes a zip with index.html at the root and needs relative asset paths (itch.io docs), and you can see why the prompt above asks for them. GitHub Pages is the other easy option, since it serves HTML, CSS and JavaScript files straight from a repository (GitHub docs).

Once it’s live, you can submit your game to us. The editorial policy says what we accept, and we review every game before it’s listed. If you’re thinking of charging for it, read selling AI-made games first.

What people have built with Google’s tools

Google’s tools show up less often in our catalog than Claude Code, Cursor and Codex. The Gemini app entries don’t say whether Canvas was used, and AI Studio and Gemini CLI appear on 0 entries.

These counts are current as of November 21, 2026. Creators name the tools themselves, in the jam submission form or a project README, and we don’t independently verify those credits (how we count).

Measure Antigravity games Gemini app games All 588 listed games
Listed games 20 10 588
Use Three.js 16 7 322
Playable on mobile 14 6 293
Online multiplayer 4 2 200
Recorded online support Share of listed games
Games crediting Antigravity, the Gemini app, AI Studio, Gemini CLI, Jules or Firebase Studio 20%
All catalog games 34%

The examples further down this page are a pretty mixed bunch, from a 3D demolition simulator and a point-and-click adventure to a neon tightrope balancer, a survival runner with a hostile AI and a Phaser brick breaker. Two of them come from Vibe Jam 2026, where the jam rules required at least 90% of the code to be written by AI.

Demolition Drive placed 74th there, and Scavender placed 81st.

You’ll find the rest on the Antigravity hub, the Gemini app hub or under games made with Gemini.

Choose a model for the code, then budget for API calls

For new projects, Google’s model list points you to Gemini 3.5 Flash-Lite or Gemini 3.8 Flash.

The newer Flash model arrived in September. Google’s announcement offers it in AI Studio, in Antigravity and, if you subscribe to Google AI Pro or Ultra, in the Gemini app. The free Gemini plan lists 3.6 Flash and varying access to 3.1 Pro.

Model Google’s description Status
Gemini 3.8 Flash “Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows” Stable
Gemini 3.5 Flash-Lite “Our fastest, most cost-effective 3.5 model for high-throughput execution” Stable
Gemini 3.1 Pro “Advanced intelligence, complex problem-solving skills, and powerful agentic and vibe coding capabilities” Preview
Nano Banana 2 “Powerful, high-efficiency image generation and editing, optimized for speed and high-volume use cases” Stable

If you want more than the free plan, Google AI Pro costs $19.99 a month and Google AI Ultra starts at $99.99 a month (plans). Building a game in Canvas, Build mode or Antigravity actually falls under those plans, not under API prices.

API prices only come in when your game itself calls the Gemini API. Flash has a free tier for that, which is handy while you’re testing, and paid API usage is billed separately (API pricing).

Model and period Input Output
Gemini 3.8 Flash, through December 31, 2026 $0.75 per million tokens $3.75 per million tokens
Gemini 3.8 Flash, after that $1.50 per million tokens $7.50 per million tokens

Images are a whole different story. Nano Banana 2 has no free API tier and costs $0.067 per 1K image (API pricing), so a game that generates art while people play pays from the first picture.

If you’re making sprites and textures ahead of time instead, game assets with AI goes through the options.

Compare Gemini with the other tools

Cursor game development and make games with ChatGPT answer the same question for two other tools.

If you’d like to see the coding agents side by side, read Claude Code vs Codex for game development and best AI for game development. Best AI game makers does the same for game makers, by export, price and rights.

New to all of this? How to make a game with AI and vibe coding a game cover the basics, whatever tool you pick. For more games to look at, there are hubs for Google games, Three.js games and the Vibe Jam 2026 archive.

Games to look at

Questions

Can Gemini make a game?

Yes, at least browser games. Google's Canvas page says Canvas turns a description into a working, shareable app or game, and its Gemini 3.8 Flash announcement shows a 3D game built in Antigravity from a simple prompt. We list 30 games whose creators name a Google tool.

Is there a free way to make a game with Gemini?

As of September 2026, yes. Canvas comes with the free Gemini plan, Google AI Studio usage is free of charge in all available regions, and Antigravity is free for individual developers with weekly limits. Deploying to Google Cloud Run is the part that can cost money.

Can I make a game in Google AI Studio?

Yes. Build mode turns a prompt into a web app or a native Android app with a live preview, and you can deploy it to Cloud Run or download it as a ZIP. Google's docs say the Antigravity Agent powers it.

What happened to Gemini CLI?

Google introduced Antigravity CLI on May 19, 2026 and encouraged Gemini CLI users to migrate. It said Gemini CLI would stop serving Google AI Pro, Ultra and free Gemini Code Assist users on June 18, 2026, while enterprise licenses and paid API keys weren't affected.

Which Gemini model should I use for games?

For code, Gemini 3.8 Flash. Google's docs tell new projects to use it or Gemini 3.5 Flash-Lite, and they describe 3.8 Flash as built for long-horizon software engineering. For art, Google lists Nano Banana 2 and Nano Banana Pro as its image models, and its own 3.8 Flash demo game used textures made with Nano Banana.