The jump in quality is hard to overstate. A year ago, asking Claude to generate a 3D scene got you flat color blocks and basic geometry. Today, Claude Opus 5 is spitting out first-person shooters, submarine games, and kart racers — complete with physics, textures, and original music — all from a single text prompt. Everything runs as a standalone HTML file in the browser, built on Three.js. No external assets, no image files, no pre-made libraries beyond the framework itself.
Community testers have been running the model through its paces with increasingly ambitious prompts. Matt Shumer posted a Call of Duty-style shooter where every texture is generated code. Alex Ermolov’s snowboard demo came back with no visual glitches and what he described as sliding physics that “feel right.” The most telling comparison came from Chetaslua, who ran the same ballista prompt through GPT-5.6 Sol and Kimi K3. Both competitors produced recognizable siege weapons, but Opus 5’s version included a full castle backdrop, a detailed control panel, and substantially more mechanical complexity.
The approach is fundamentally different from what Google is doing with Genie 3 or the video models that get called world simulators. Those systems render frame-by-frame and hand you back a video stream — watchable but not editable. Opus 5 outputs actual code: procedural geometry, shader-based textures, and runtime physics that you can open up and modify. It’s a distinction with real practical weight. An animator or game designer can take the HTML file and iterate, rather than starting from scratch.
Minecraft clones have become the unofficial community benchmark, and the results are genuinely impressive. Pankaj Kumar’s version features an infinite procedural world with 15 biomes and both survival and creative modes — a build that consumed 25 million tokens. Atomic.chat ran a head-to-head physics stress test across four models, asking each to simulate a tornado tearing through a field and a wrecking ball swinging into a building. Only Opus 5 handled all three scenes convincingly. This isn’t a rigorous benchmark by any stretch, but when anyone can recreate the prompts and see the results on screen, the comparisons carry more weight than a benchmark score ever could.
