We asked four AI coding agents to rebuild Minesweeper–the results were explosive (arstechnica.com)

In a recent test, Ars Technica evaluated the coding capabilities of four prominent AI models—OpenAI’s Codex (GPT-5), Anthropic’s Claude Code (Opus 4.5), Google’s Gemini CLI, and Mistral Vibe—by challenging them to recreate the classic Windows game Minesweeper. Each AI agent was tasked with not only replicating the original game but also implementing an innovative gameplay feature and ensuring mobile touchscreen support. This test, conducted without prior notice or privileged access from the companies, aimed to assess how effectively these models could handle a straightforward coding project, highlighting the current capabilities and limitations of AI in software development. The findings reflect the ongoing debate within the AI/ML community regarding the reliability of coding agents. While some developers are skeptical about the potential for trust-worthy AI-generated code due to past inaccuracies, the results suggest that these frontier models are improving. The competition among AI models to deliver functional code without human intervention offers valuable insights into their development trajectories. However, it also underscores a critical reality: even as AI evolves, human oversight remains essential for debugging complex outputs, as the use of unmodified code in this test inevitably revealed inconsistencies and inefficiencies.