Key takeaways

  • A first-person shooter, a submarine game, a kart racer, a Minecraft clone: over the past few days, people have generated all of these from…
  • Writing a scene from scratch takes more than clean syntax because the model has to know how the world actually looks and moves, how a…
  • Matt Shumer posted a Call of Duty-style shooter where everything on screen, right down to the textures, is generated code.

What happened

A first-person shooter, a submarine game, a kart racer, a Minecraft clone: over the past few days, people have generated all of these from a single prompt each. The models write the geometry, the textures, and sometimes even the music as code that runs straight in the browser.

The Opus 5 demos work differently: Geometry is procedural, built by code that computes points and surfaces at runtime. Textures exist as shaders, which is to say more code. js, one HTML file is enough. Anthropic says in its Opus 5 announcement that the model produces "much stronger visual outputs" than earlier versions.

One testing partner quoted there calls the results the best animations, games, and 3D work he's seen from any Opus model. It's a different approach from Google's Genie 3 or the video models sometimes described as world models. Those systems compute every frame and hand back an image sequence. The output only exists as a video stream, so there's nothing to edit or reuse afterward.

What Opus 5 produces is code you can open and change. Beyond the one-off demos, the community has settled on an informal test of its own: rebuild Minecraft from a single prompt. Pankaj Kumar's Opus 5 version has an infinite procedural world, 15 biomes, and both survival and creative modes. Every graphic asset comes from code, he writes, with textures and sounds generated at runtime.

Why it matters

Writing a scene from scratch takes more than clean syntax because the model has to know how the world actually looks and moves, how a snowboard slides over snow, or how wind travels through a field of grass. None of that is in the prompt; it has to come from the model itself.

Matt Shumer posted a Call of Duty-style shooter where everything on screen, right down to the textures, is generated code. "Not a single external asset was used," he wrote on X. The prompt and source code are up on GitHub under the name Claude-of-Duty. Alex Ermolov put out a snowboard demo and said Opus 5 is on par with Fable and ahead of every other model he's tested.

" Chetaslua's ballista makes the gap to other models pretty stark. The Roman siege weapon sits in front of a castle, has a complete control panel, and feels closer to a finished game than a prototype. 6 Sol and Kimi K3. Both managed a recognizable ballista, but with fewer working parts and far less going on around it. User Chris ran a comparison that shows how quickly this has improved.

He put year-old output from Claude 4 Opus next to what Opus 5 produces today. For the new run, he asked for a car on a dirt road, explicitly without textures. Opus 5 came back with mud tracks, vegetation, and layered lighting. A year ago, the same kind of request produced flat blocks of color.

Lentils packed a scenic landscape into one HTML file, with millions of grass blades bending to simulated wind. Pietro Schirano built a submarine game where the model generated the 3D objects, textures, and music itself. Ryan Campbell put his kart racer online as a playable website. In traditional 3D development, developers assemble assets from libraries, load textures as image files, and handle physics through an engine.

What to watch

The build ate 25 million tokens. Other head-to-head tests follow the same script. js and lined up the results. chat went a step further and had all four models build three physics scenes, among them a tornado tearing across a field and a wrecking ball swinging into an apartment block. chat's account, only Opus 5 handled all three convincingly.

Prompts like these work as a rough first gauge, since anyone can recreate them and judge the results on screen. Programming skills aren't required, though they can help with the prompting. A benchmark in any methodical sense, however, this is not. There are no standardized tasks, no rating scale, and no control over how many attempts each model gets.