Vibe coding games: why October 2026 is an inflection point
Simon Willison argues that October 2026 is when AI models became good enough at game design to make vibe coded games genuinely fun. He built four browser games mostly with Claude Opus 5.5 and shares the prompts and workflow he used. The post is a practical look at iterating with coding agents, and at where GPT-6 models still fall short.
Key points
- Start a game prototype in a chat interface like claude.ai or chatgpt.com with a single prompt.
- Asking explicitly for three.js worked well for Willison's 3D browser games.
- Iterate on fun with follow-up prompts such as 'Make it a LOT more fun' and fix gameplay issues one by one.
- Move to Codex or Claude Code once the prototype proves itself.
- Willison reports that Claude Opus 5.5 outperformed the GPT-6 models he tried on game design and fun.
Simon Willison argues that October 2026 marked a turning point for vibe coding games, because current models now handle game design, not just code and assets. He built four browser games in HTML and JavaScript, mostly by prompting Claude Opus 5.5 with little guidance. A simple request like 'Make it a LOT more fun' now produces noticeably better gameplay.
Some projects were built in one shot, while Battle Drones, a multiplayer real-time strategy game, took many iterations and a move to three.js. Willison finds that GPT-6 Astra and GPT-6 Pro can build games but lack Opus 5.5's sense of fun. His advice: prototype in a chat interface with a mobile-friendly three.js prompt, then switch to Codex or Claude Code once the idea holds up.
Real-world use
- A developer building a mobile-friendly browser game prototype over a weekend.
- A small team testing game concepts and gameplay loops before committing to a full build.
- An educator producing a quick interactive 3D demo for a class using a chat model.
Why it matters
Prototyping a playable game now takes a chat session and a few iterations, so the real work shifts to game concepts, playtesting and disciplined follow-up prompts.