Claude Opus 5 pushes prompt-to-game AI from rough color blocks to full 3D prototypes with physics and music

In a remarkable leap forward for artificial intelligence in gaming, Anthropic’s Claude Opus 5 has demonstrated the ability to generate fully functional 3D game worlds from a single text prompt. This innovative model can create everything from first-person shooters to Minecraft-style environments without relying on any external files or pre-made assets. The model autonomously writes all necessary code for geometry, textures, physics, and even music, rendering the output as an editable HTML file that runs directly in the browser.

Six demos built without a single external asset

Recent community tests have showcased the impressive capabilities of Opus 5, with users generating a variety of games using nothing but a single prompt. For instance, Matt Shumer crafted a Call of Duty-style shooter entirely from code generated by Opus 5, proudly noting on X that “not a single external asset was used.” The prompt and source code for this project, dubbed Claude-of-Duty, are available on GitHub.

Matt Shumer’s first-person shooter was built entirely by Opus 5, with zero external assets. | Image: Matt Shumer

Another user, Alex Ermolov, created a snowboard demo that he claims rivals the quality of established games like Fable, stating that Opus 5 produced a flawless experience with realistic sliding physics. Chetaslua’s ballista project further highlighted the advancements, showcasing a Roman siege weapon with a complete control panel that felt more like a finished game than a mere prototype. Previous attempts using other models, such as GPT-5.6 Sol and Kimi K3, yielded less complex results.

<video controls width="640" height="360" src="https://the-decoder.com/wp-content/uploads/2026/08/opus5snowboard_demo.mp4″>

In a side-by-side comparison, user Chris demonstrated the significant progress made by Opus 5 over its predecessor, Claude 4 Opus. By requesting a simple scene of a car on a dirt road without textures, Opus 5 produced a rich environment complete with mud tracks and layered lighting, whereas the earlier version resulted in flat blocks of color.

Claude 4 Opus from last year (bottom) compared to Opus 5 (top). | Image: ChrisGPT

Other demos have further illustrated the model’s versatility, with Lentils creating a scenic landscape featuring millions of grass blades swaying in the wind, and Pietro Schirano developing a submarine game where all 3D objects, textures, and music were generated by the model itself. Ryan Campbell even launched a playable kart racer online, showcasing the potential for interactive experiences.

Procedural code replaces traditional asset pipelines

Unlike traditional 3D development, which relies on assembling assets from libraries and loading textures as image files, the Opus 5 approach is fundamentally different. It utilizes procedural code to construct geometry, generating points and surfaces in real-time. Textures are created as shaders, meaning they are also coded rather than pre-made. This allows for a streamlined process where all elements, including physics and controls, are integrated into a single HTML file that runs seamlessly in the browser.

Anthropic has emphasized that Opus 5 produces significantly stronger visual outputs compared to earlier models, with one testing partner describing the results as the best animations, games, and 3D work he has encountered from any Opus iteration.

This innovative method contrasts sharply with systems like Google’s Genie 3 or video models that produce image sequences without editable outputs. Opus 5, on the other hand, generates reusable code that can be modified and expanded upon by developers.

Minecraft clones are becoming the community’s go-to stress test

The community has informally adopted a benchmark by attempting to recreate Minecraft from a single prompt. Pankaj Kumar’s version using Opus 5 features an infinite procedural world with 15 biomes and both survival and creative modes, all generated in real-time. This ambitious project consumed 25 million tokens.

Other comparative tests have followed suit, with users challenging Opus 5 against models like Fable 5 and GPT-5.6 Sol to generate complex 3D scenes. One notable test involved creating a tornado simulation, where only Opus 5 convincingly handled the physics of all three scenarios presented.

Visual comparison of four AI models running a tornado simulation from the same prompt. | Image: atomic.chat

These prompts serve as a preliminary gauge of performance, allowing anyone to recreate them and assess the results visually. While programming skills can enhance the prompting process, they are not strictly necessary. However, it is important to note that these tests lack standardized tasks and rating scales, making them more of an exploratory exercise than a rigorous evaluation.

AppWizard