Anthropic Models Decompile a Game and Port It to the Browser Overnight

Decompiling a commercial game and moving it to the browser used to take weeks, even months, of work by people who know assembly, legacy graphics APIs, and resource formats. Now, according to 3DNews, it takes one night. Almost any user of Anthropic's frontier models can start such a process. The resulting ports open directly in a browser tab, work fairly well, and in some respects surpass the originals.

The article names no specific projects and gives no statistics or measurement methodology. Its illustration comes from Radical Entertainment, the studio known for Prototype and The Simpsons: Hit & Run. So what follows covers what can be asserted from the claim itself, and what remains unknown.

What such a pipeline looks like

Classic reverse engineering has several layers. First, the binary is disassembled, which yields hard-to-read pseudocode. Next, data structures, function names, and game logic are reconstructed from it. After that, the rendering and audio layers are rewritten for modern APIs. Finally, everything is packaged into a form the browser understands: via WebAssembly, or by rewriting in JavaScript and TypeScript.

The most labor-intensive part is routine. You have to read thousands of functions, give them meaningful names, and match their behavior against the original. For language models, especially agentic ones, this is a nearly ideal task: lots of repetitive work, a clear pass/fail check ("does it run or not"), and the option to work in loops with automated testing. That is probably why the timeline collapsed to a night. The report does not describe the port authors' exact workflow, so this is a plausible reconstruction, not a retelling.

Why a port can be better than the original

The claim of superiority over the source sounds bold, but it is easy to explain. A browser port usually sheds its dependence on old hardware and a 640×480 resolution. It gains widescreen output, a stable frame rate, and modern controls. It needs no installation and no compatibility with a current version of Windows. Bugs that only appeared on newer systems also go away.

Behind these improvements there may be losses: inaccurate physics, audio discrepancies, broken scripts in rare scenes. There is no data on content coverage or on how closely the port matches the original.

What the report tells us

• Timeline: roughly one night instead of weeks and months of manual work.
• Tool: Anthropic's frontier models.
• Barrier to entry: no professional reverse engineer needed.
• Result: the game runs in the browser, with quality described as decent and in places better than the original.
• Not disclosed: the list of games, token consumption and cost, the amount of manual fixes, the legal status of the projects.

Legal side and practical takeaways

The main question lies outside the technology. A game's decompiled code and resources remain the intellectual property of the rights holders. The community has long drawn a line: you can publish tools and code that reproduce an engine, but you cannot distribute the original models, textures, and sound. If a port is built from files the user supplied, the situation looks milder. If the assets sit on a public server, rights holders have a direct basis for claims. When anyone can make such ports in one night, the number of disputed cases will rise sharply.

For publishers, the takeaway is simple. Archives of old games that sat unsupported for years can now be revived by outside enthusiasts. Publishers should settle on a policy in advance: ban, tolerate, or release official web versions themselves. For developers, this is a gauge of what agentic tools can do with atypical, poorly documented codebases. Anyone who wants to repeat the experiment should stay on the safe side: use games whose rights you own or that are published freely, and do not distribute other people's assets.

For now this is more a loud signal than a measured result. To judge the scale, we need open repositories, comparative tests, and the rights holders' position.