Claude Opus 5 is turning a single text prompt into browser-based 3D game prototypes that look and behave far more complete than the rough visual experiments associated with earlier systems. Recent community demos include a first-person shooter, a submarine game, a kart racer, a Minecraft clone, and physics-heavy scenes generated without external files or pre-made assets.
What the demos show
The most striking claim across the examples is not just that Claude Opus 5 can write working code. It is that the model can create the pieces normally supplied by a broader production workflow: geometry, textures, controls, physics, and sometimes music.
Matt Shumer shared a Call of Duty-style shooter in which the visible scene, including textures, was generated as code. He wrote, "Not a single external asset was used," and the prompt and source code are available on GitHub under the name Claude-of-Duty.
Alex Ermolov released a snowboard demo and described Opus 5 as on par with Fable and ahead of every other model he had tested. According to him, the first run had no visual glitches, and the "sliding physics feel right."
Other examples broaden the range. Chetaslua created a ballista scene with a Roman siege weapon in front of a castle and a complete control panel. Lentils generated a scenic landscape in one HTML file with millions of grass blades bending to simulated wind. Pietro Schirano built a submarine game where the model generated 3D objects, textures, and music itself. Ryan Campbell published a kart racer as a playable website.
Why this is different from a normal asset pipeline
Traditional 3D development usually depends on asset libraries, image files for textures, and engines to handle physics. The Claude Opus 5 demos described here use a different route: the model writes procedural code that creates the scene at runtime.
In that approach, geometry is calculated rather than imported. Textures are expressed as shaders, meaning they are also code. The same generated file can include physics, controls, and rendering logic.
Because the output runs in the browser on Three.js, one HTML file can contain the full prototype. That matters because the result is editable code, not just a rendered clip. A developer can open the file, inspect how it works, and change it afterward.
The source article contrasts this with Google's Genie 3 and video models sometimes described as world models. Those systems generate frames and return an image sequence. Claude Opus 5, in these examples, produces code that can be reused and modified.
The comparison with earlier models
Several tests described in the source point to a visible jump over prior model output. User Chris compared year-old output from Claude 4 Opus with an Opus 5 result today. The new prompt asked for a car on a dirt road, explicitly without textures.
Claude Opus 5 returned a scene with mud tracks, vegetation, and layered lighting. The older Claude 4 Opus example produced flat blocks of color. That contrast is central to the broader claim: the newer model appears better at filling in visual structure and environmental detail from sparse instructions.
Chetaslua's ballista comparison also showed a gap. GPT-5.6 Sol and Kimi K3 both created a recognizable ballista, but the source says they had fewer working parts and much less activity around the scene. The Opus 5 version looked closer to a finished game than a prototype.
Anthropic also says in its Opus 5 announcement that the model produces "much stronger visual outputs" than earlier versions. A testing partner quoted there described the results as the best animations, games, and 3D work he had seen from any Opus model.
Minecraft has become a stress test
Beyond isolated demos, the community has started using a specific challenge as an informal measure: rebuild Minecraft from a single prompt. Pankaj Kumar's Opus 5 version includes an infinite procedural world, 15 biomes, and both survival and creative modes.
According to Kumar, every graphic asset comes from code, with textures and sounds generated at runtime. The build used 25 million tokens.
Other head-to-head tests use similar prompts across multiple systems. Harshith asked Opus 5, Fable 5, GPT-5.6 Sol, and several other models to generate a 3D Airbus H145 in Three.js, then compared the outputs. atomic.chat tested four models on three physics scenes, including a tornado crossing a field and a wrecking ball hitting an apartment block. By atomic.chat's account, only Opus 5 handled all three convincingly.
Useful evidence, but not a benchmark
These demos are useful because the outputs are visible and, in many cases, playable. People do not need programming skills to compare whether a scene looks coherent, whether motion seems plausible, or whether the controls and objects appear to work together.
That does not make the results a formal benchmark. The source notes that there are no standardized tasks, no rating scale, and no control over how many attempts each model receives. A polished public demo may show what is possible, but it does not prove consistent performance across every prompt.
Still, the pattern is important. Claude Opus 5 is being tested not just on code correctness, but on whether it can infer how a world should look and move when the prompt leaves most of the details unstated. In these public examples, the model is not only assembling syntax. It is generating interactive 3D scenes with procedural geometry, shader-based textures, browser rendering, physics behavior, and sometimes audio, all from text.