Modern Claudefare: A Full FPS Built Entirely With Claude Opus 5

A developer working under the name 0xRishi has released Modern Claudefare, a first-person shooter game generated almost entirely by Anthropic's Claude Opus 5 running in high mode. The project totals 84,100 lines of code and includes remade versions of four classic multiplayer maps from the shooter genre, playable either solo against bots or online with other players. The build reportedly used techniques drawn from Matt Shumer's Gauntlet Loop approach, a method for structuring long-running AI coding sessions so that a model can work through a large, complex codebase in stages rather than in a single pass. This kind of iterative loop has become increasingly common among developers pushing frontier models to handle full application builds, letting the AI check its own output, catch errors, and continue building on previous work without a human rewriting large sections by hand. What makes Modern Claudefare notable is not just the scale of the code but the genre. First-person shooters require precise handling of physics, networking, hit detection, map geometry and rendering performance — areas where small bugs are highly visible and gameplay-breaking. Building a multiplayer-capable FPS with this much code, largely through AI-generated output, is a stronger test of a model's coding ability than many of the web apps or utility scripts typically used to showcase coding models. The release fits into a broader pattern of developers using top-tier coding models to attempt increasingly ambitious software projects as a way of benchmarking real-world capability. Games, in particular, have become a popular test case because they combine multiple demanding subsystems — real-time rendering, input handling, networked state synchronisation and asset management — into a single project. A model capable of producing a playable, online-ready shooter at this scale suggests meaningful gains in reliability over long, complex coding sessions, though as with all such demonstrations, questions remain about how much manual correction, debugging or guidance the developer provided along the way. Anthropic has not commented directly on the project, and full technical details of the development process, including how much of the 84,100 lines were generated without human edits, have not been independently verified. Still, the project adds to a growing body of examples suggesting that current frontier coding models are capable of sustained, large-scale software builds when paired with structured prompting techniques like the Gauntlet Loop.