Claude Opus 5.5 builds a browser Spider-Man game, with Shikhar steering

Why it matters

Spiderbench is a concrete test of AI-assisted production across code, 3D assets, procedural environments and rendering. Shikhar's own description sets the limit: it is a human-directed benchmark, not a finished commercial game, and he has not calculated the model's API-equivalent cost.

A browser-based Spider-Man fan game built with Claude Opus 5.5 has become a public test of how far an AI coding agent can take a real-time 3D project when a person directs the work. Developer Shikhar (@xikhar) posted the third iteration of Spiderbench on September 27th, describing a workflow that combined Blender, generated images and Three.js. He said he began with medium reasoning to save tokens and had spent three days using a Claude 5x subscription; he had not checked what the work would have cost through the API. His post on X calls the project a benchmark, not an attempt to make a commercial game.

The Spiderbench repository describes a non-commercial technical demonstration made with Claude Code. Its README says Claude produced the code, shaders, procedural city, Blender-built models and generated textures under human direction, with Shikhar steering through feedback and reference images. That division of labor is central to what the demo shows: the model handled a broad set of connected production tasks, while a person chose the target, supplied references and iterated on the result.

The project goes beyond a static scene. Its README describes third-person web-swinging, wall-running, perching, zipping and diving, controlled by a custom physics and animation state machine. The game world is a procedural Manhattan-style island, with an authored street network and thousands of buildings, rooftops, parks, traffic and pedestrians. Its rendering pipeline uses Three.js and WebGL2, with shadows, screen-space lighting, ambient occlusion, reflections, bloom, temporal anti-aliasing and motion blur. Blender Python scripts generate models and animation clips; the project also uses AI-generated textures and advertising art.

Those details make Spiderbench a substantial software demo, but they do not establish that an AI agent can independently produce a finished game. Shikhar said the project was not meant to ship, and described it as an effort to benchmark the model. The repository likewise presents it as a technical demonstration. Its public materials show a specific result assembled through multiple iterations and human feedback, rather than a standardized head-to-head test of game-development systems.

The timing put the demo in the first week of Claude Opus 5.5's release. Anthropic announced the model on September 22nd, positioning it as an agentic coding model and reporting improvements on its own coding evaluations. The company lists API pricing at $4 per million input tokens and $20 per million output tokens. Shikhar's account offers no equivalent cost estimate for Spiderbench: his three-day subscription use cannot be translated into an API bill from the information he shared.

Spiderbench is available as source code and assets on GitHub. The repository says it needs Node.js 20.19 or newer (or Node.js 22.12 or newer), a WebGL2-capable browser and recommends a discrete GPU. It provides commands to install dependencies, run a local development server and build the project. Shikhar said the game ran on both an M4 Mac and a laptop with an Nvidia 3050, while also noting that the goal was not to release it as a product.

The repository identifies Spiderbench as an unofficial fan project and says it is not affiliated with or endorsed by Marvel, Disney, Sony or Insomniac Games. It says the project is not for sale and may not be redistributed or used commercially. That scope matters: Spiderbench's evidence is about how an AI-assisted workflow can produce an ambitious, locally runnable prototype. It does not show a route to commercial distribution or replace the work required to turn a prototype into a supported game.