People are building playable 3D games from a single prompt, and the results stopped looking like blocks
A first-person shooter, a kart racer, a submarine game and a Minecraft clone, each generated by Claude Opus 5 from one prompt with no external art files. Everything on screen is code that runs in the browser.
Over the past week a small genre has taken over parts of the AI internet: ask a model for a 3D game, get a playable one back. Not a mockup, not a video, an HTML file you can open in a browser and control. The demos come from Claude Opus 5, and what makes them notable is not that they exist but that not a single external file went into them.
Matt Shumer posted a Call of Duty style shooter where every texture on screen is generated code, and put the prompt and source on GitHub. Alex Ermolov built a snowboard demo and said the sliding physics felt right on the first run, with no visual glitches. Pietro Schirano’s submarine game has the model generating the 3D objects, textures and music. Ryan Campbell put a kart racer online as a playable website. Someone called Lentils packed a landscape with millions of grass blades bending in simulated wind into one file.
Rebuilding Minecraft from a single prompt has become the community’s informal stress test. Pankaj Kumar’s version has an infinite procedurally generated world, 15 biomes, and both survival and creative modes, with every texture and sound produced at runtime. It consumed 25 million tokens to build.
The clearest before and after came from a user called Chris, who ran the same request through last year’s Claude 4 Opus and this year’s Opus 5. He asked for a car on a dirt road, explicitly without textures. A year ago that produced flat blocks of colour. Opus 5 returned mud tracks, vegetation and layered lighting.
What is behind this. Traditional 3D development assembles a game from parts: art files from a library, textures loaded as images, physics handled by an engine. These demos have none of that, which is why they are interesting. The geometry is procedural, meaning code calculates the points and surfaces while the scene runs. The textures are shaders, which is to say more code. The physics is written from scratch. All of it runs on Three.js in a browser, so one file is enough. That shifts the difficulty: the model has to know how the world actually looks and behaves, how a board slides over snow or how wind moves through grass, because none of that is in the prompt. It is also a different thing from Google’s Genie 3 or the video-based world models, which compute each frame and hand back a stream you cannot edit. What Opus 5 produces is code you can open and change.
A fair caveat: these are not benchmarks. There are no standard tasks, no scoring, and no control over how many attempts each model got before someone posted the good one. Anthropic’s own launch material claims stronger visual output, which is exactly the claim these demos are being used to support.
What this means for you: if you have ever wanted to make a small game and bounced off the tooling, the barrier just dropped a long way, and you do not need programming skills to try it, though they help with the prompting. Ask for something specific and small, a physics toy or a single mechanic, rather than a whole game. For everyone else, treat this as a decent proxy for a harder question: how much does a model know about how the physical world behaves when nobody tells it?
Sources
AI is finding a flood of security flaws. Almost none of them get attacked
VulnCheck counted 1,061 vulnerabilities found with AI help in the first half of 2026. Fourteen were confirmed exploited, a rate of 1.3 percent, which is roughly the same as vulnerabilities found any other way.
