Claude Opus 5: From Color Blocks to Full 3D Games

Claude 4 Opus turned a game prompt into flat blocks of color a year ago. This week, developers ran similar prompts through Claude Opus 5 and got mud tracks, wind-blown grass fields, working siege weapons, and a Minecraft clone with its own soundtrack — all generated from single prompts, all rendered as one editable file.

What Happened

Over the past few days, a string of developers independently prompted Claude Opus 5 to build playable 3D games and posted the results, as compiled by The Decoder. Matt Shumer built a Call of Duty-style first-person shooter and said “not a single external asset was used” — every texture, model, and sound came out of the model’s own code. Alex Ermolov’s snowboard demo had sliding physics that “feel right,” with no visual glitches on the first run. Chetaslua built a ballista siege weapon complete with a working control panel, then ran the identical prompt through rivals GPT-5.6 Sol and Kimi K3: both returned a recognizable ballista, but with fewer working parts and far less detail surrounding it. Pietro Schirano’s submarine game and Pankaj Kumar’s Minecraft clone both leaned on the model’s ability to generate not just geometry and shader-based textures but music, written as code that renders live in the browser via Three.js.

Why It Matters

The jump is visible in direct before/after tests. One tester, going by Chris, gave Claude 4 Opus and Opus 5 the identical prompt a year apart — a car on a dirt road, explicitly without textures. Claude 4 Opus returned flat color blocks; Opus 5 returned mud tracks, vegetation, and layered lighting, generated procedurally rather than pulled from an asset library. That’s the mechanism behind every demo in this batch: the model isn’t retrieving pre-built game assets, it’s writing the geometry, the shaders, the physics, and sometimes the music as code, then rendering the result live in a single HTML file anyone can open, edit, and reuse. Ermolov, who tested the same prompt across models, said Opus 5 is on par with Anthropic’s own Fable line and ahead of every other model he’s tried — which, paired with the ballista result against GPT-5.6 Sol and Kimi K3, points to a real gap between Opus 5 and its closest frontier competitors on this specific task. It also tracks with what Stack Archive covered when Opus 5 posted a 0% success rate against browser-based prompt injection attacks earlier this cycle — a model turning in unusually strong, independently verified results across more than one evaluation.

What This Means for Builders

None of this is a rigorous benchmark, and The Decoder is explicit about that: there are no standardized tasks, no rating scale, and no control over how many attempts each tester got before landing on a result worth posting. But the token cost is a real number, not an anecdote — Kumar’s Minecraft clone alone burned 25 million tokens, which puts a price tag on “one prompt, one game” that’s easy to skip past in a demo reel. For anyone evaluating Opus 5 for rapid prototyping, game jams, or internal tooling demos, the capability jump from Claude 4 Opus is real and independently replicated across more than half a dozen testers — but budget for token spend the way you would for compute, not for a single API call.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.