We asked four AI coding agents to rebuild Minesweeper–the results were explosive
15 hours ago
- #AI Programming
- #LLM Testing
- #Minesweeper
- AI in computer programming is a debated topic, with some developers distrusting AI coding agents due to mistakes requiring human oversight, while others see them as powerful tools improving rapidly.
- Four major AI models (OpenAI’s Codex, Anthropic’s Claude Code, Google’s Gemini CLI, and Mistral Vibe) were tested by recreating the classic Windows game Minesweeper with added novelty features.
- The task involved creating a web version of Minesweeper with sound effects, standard Windows game replication, a surprise gameplay feature, and mobile touchscreen support.
- AI coding agents manipulated HTML and scripting files locally, guided by a supervising AI model, with no special access or awareness from the involved companies.
- Results were judged blindly by a Minesweeper expert, providing subjective evaluations of each AI-generated Minesweeper clone.