====== # Bot-generated -- edit freely. ======
Claude Fable 5.2 Creates Stunning Underwater World In Just 97 KB, Runs At 180 FPS On An RTX 5070 Ti Without A Game Engine [wccftech.com]:
Anthropic's Claude Fable 5.2 continues to showcase its increasingly impressive coding capabilities, this time generating an entire interactive underwater environment that fits inside a single 97 KB HTML file , without relying on traditional game engines, textures, 3D models, or external assets.
The project, appropriately named Abyssal Garden , was shared on Reddit by developer Noniv , who used Claude Code to create a surprisingly beautiful underwater world featuring dense kelp forests, sunken ships, manta rays, schools of fish, and atmospheric lighting.
Even more remarkably, the entire experience runs directly inside a web browser using WebGPU [webgpu.org], and the developer reports performance reaching approximately 180 FPS at 1440p on an NVIDIA GeForce RTX 5070 Ti .
An Entire Underwater World Generated Procedurally, Without Traditional Assets
What makes Abyssal Garden particularly impressive is its tiny footprint. Rather than loading conventional textures, meshes, or audio files, everything is procedurally generated from code when the environment launches .
According to the project's GitHub repository [github.com], Claude generated the WGSL shaders, GPU-sculpted seabed, approximately two million kelp leaves and plants, tens of thousands of rocks, and thousands of underwater creatures and decorative objects .
The environment also features dynamic sunlight filtering through the ocean surface, animated caustics, volumetric-looking light rays, floating particles, and an atmospheric soundtrack synthesized in real time through the Web Audio API .
Remarkably, the developer claims the initial version originated from a single detailed prompt [github.com]submitted to Claude Code. The instructions specified visual quality, procedural generation, stable frame pacing, and a self-contained HTML deliverable , explicitly prohibiting engines such as Three.js and Babylon.js .
The prompt also established an interesting verification procedure: a separate AI subagent would examine rendered screenshots and identify visual defects, while GPU timing measurements would evaluate performance during a scripted 60-second fly-through.
However, it wasn't an entirely one-shot creation. Subsequent prompts introduced collision detection, a start menu, and adjustable graphics settings . The developer estimates approximately two to three hours of actual Claude processing , spread over roughly three days because of subscription usage limits.
Surprisingly Impressive Performance, And You Can Try It Yourself
Abyssal Garden reportedly achieves around 180 FPS at 2560 × 1440 on an RTX 5070 Ti , although these performance figures come from the developer and haven't been independently verified.
The experience also supports less powerful systems through Low, Medium, and High detail presets. According to the developer, even a modest Samsung Galaxy A16 smartphone can run the simulation at approximately 20 FPS using Low settings and a reduced resolution .
Visitors can freely explore the environment using WASD controls, mouse movement, and vertical swimming controls . The project is available to play directly in your browser [github.io], preferably using Chrome or Edge with WebGPU support . Its source code is also freely available under the MIT license .
This latest experiment follows another ambitious Claude Fable 5.2 demonstration that we previously covered [wccftech.com], involving an Unreal Engine 5-based Gears of War: E-Day [wccftech.com]-inspired shooter generated through extensive autonomous AI coding .
Of course, Abyssal Garden is an interactive graphical demonstration rather than a complete video game, and procedural rendering is hardly a new concept. Nevertheless, creating such a visually rich experience with a remarkably small amount of code offers another compelling example of how far AI-assisted software development has progressed.
News Source: Noniv on Reddit [reddit.com]
See also: