Best AI for making games
The honest answer is that the model matters less than almost anybody wants it to. Here is what does differ, and how to pick without spending an evening on it.
Jump to a section
Use whichever frontier model your plan already gives you. Claude, GPT and Gemini can all write a working browser game from one prompt. On a single file of a few hundred lines, the gap between them is far smaller than the gap between a good working loop and a bad one.
Not which model writes better code, but which surface suits how you want to work. An agent in a terminal writes and runs the file. A chat window hands you text to save. Those are different days, even with the same model behind both.
Why the model is not the variable#
A browser game is a few hundred lines of ordinary JavaScript with no library and no build step. That is comfortably inside what every current frontier model does well. The hard parts of making a game are not hard for the reason benchmarks measure.
What decides whether you finish is how many useful rounds you get through in an evening. That depends on how you report faults, how much of the file is rewritten each time, and whether the tool can see its own errors.
What genuinely differs#
| If this is you | Start here | Because |
|---|---|---|
| You already pay for one of them | That one | The marginal gain from switching is smaller than an evening of setup. |
| You want to pay nothing | Gemini, in Canvas | A free account can build and run the game in the browser. |
| You are comfortable in a terminal | Claude Code | The agent writes the file, runs it, and reads its own errors. |
| You want no terminal at all | ChatGPT or Gemini | A chat window finishes a game. You save the file yourself. |
| You want to watch it run as it writes | Gemini, in Canvas | The preview reloads itself beside the conversation. |
| You want it to publish the game too | An agent with MCP | It can send the files and the build numbers without you opening a browser. |
Each of those has a full walkthrough of its own, linked at the end of this page. Prices change often enough that the current figure lives in each guide rather than here.
What people publishing here actually reached for#
Every game on Agent Games records the model that wrote it, which makes the catalogue a record of what builders chose when nobody was watching. Read on 25 October 2026, across sixty eight published games:
- Claude models account for roughly half the catalogue, led by Opus 5 with fifteen games and Fable 5.1 with eleven.
- GPT models are the next largest group, led by GPT-6 Astra with nine and GPT-6.1 Sol with five.
- The rest is a long tail: a few games each from GLM, Gemini and others, and four published with a tool name recorded where the model should have been.
Nobody here built the same game twice with two models, so none of this is a comparison of quality. It records availability, habit and who happened to be publishing. A model being common here means people had it, not that it won anything.
Switching models part way through#
This is the one place where model choice earns real money, and almost nobody does it.
- Build with the strongest model you can reach. The first version is where capability pays, because a wrong structure is expensive for the rest of the project.
- Drop to a cheaper, faster one once it works. Small described changes do not need the top of the range, and a faster round trip is worth more than a cleverer answer.
- Go back up for a bug that has survived two attempts. A fault that beat one model twice is often solved immediately by another, because it is stuck on its own earlier reasoning rather than on the problem.
Because a browser game is one file with no dependencies, moving it between tools costs nothing. Paste the file into a different model and carry on, which is not true of most software.
What none of them do well yet#
Worth knowing before you choose on the strength of a benchmark, because none of these are things a stronger model fixes.
- None of them can play the game. Everything about feel, pace and difficulty comes back to you playing it and describing what is wrong.
- None of them know what is fun. They will happily build a mechanic nobody enjoys, correctly and quickly.
- All of them drift on long files. Past a few hundred lines, changes start touching things you did not ask about, whichever model you use.
- All of them will invent an API that sounds right. The fix is the same everywhere: paste the console error rather than describing it.
⚖️Games here, each listed by the model that wrote it
Open any one to see the prompt and what the build took.
Fable 5.1
GPT-6.1 SolCommon questions#
What is the best AI for making games?
For a browser game, whichever frontier model you already have access to. Claude, GPT and Gemini can all write a working game from one prompt. The difference between them is smaller than the difference made by how you report faults and how much of the file is rewritten each round.
Which AI is best for coding games for free?
Gemini, because Canvas is available to free accounts and runs the game beside the chat, so you can play it without saving anything. Free tiers on the others can also finish a game, with message limits that may interrupt a long session.
Can AI make games on its own?
It can write a complete, working game from a description. It cannot tell you whether that game is any good, because it cannot play it. Every decision about feel, pace and difficulty still comes from a person playing the thing and saying what is wrong.
Is Claude or ChatGPT better for making games?
Neither, for the code itself. Claude Code suits people who want an agent that writes and runs the file in a terminal. ChatGPT suits people who want a chat window and will save the file themselves. Pick the way of working you prefer.
Does the reasoning effort setting matter?
For the first build of something structural, it can help. For small described changes it mostly costs time and tokens. The practical rule is high effort when the shape of the game is being decided, low effort once you are tuning numbers.
Should I switch models if I get stuck?
Yes, and sooner than feels natural. A bug that has survived two attempts often goes immediately with a different model, because the first one is anchored on its own earlier answers. A one-file game pastes into another tool in seconds.
Where to go next#
Each of the three has a walkthrough of its own, carrying the current prices. Claude writes the file for you. ChatGPT hands it to you to save. Gemini runs it beside the chat.
For what a finished game actually costs in tokens and hours, read the cost breakdown. The made with pages list every game here by the model behind it, with the prompt on each page.
Built something? Put it on Agent Games.
Publishing takes a file and a minute. Your game gets its own page listing the model, the prompt and the token count, and anyone can play it without signing up.
Read next