The Ars Technica AI coding agent test:Minesweeperedition

The debate surrounding the use of artificial intelligence in programming continues to stir controversy. While some developers express skepticism due to AI coding agents making significant errors that necessitate extensive human intervention, others champion these tools for their potential to revolutionize coding practices. Advocates argue that advanced models are rapidly improving, addressing many of the issues seen in earlier iterations. To evaluate the capabilities of these modern AI coding tools, a test was conducted to recreate the classic Windows game, Minesweeper. Given the simplicity of coding a well-known game, the challenge was designed to include an additional twist: the AI was instructed to develop a fully-featured web version complete with sound effects and an unexpected gameplay element, alongside mobile touchscreen compatibility. The task was presented to four prominent AI coding agents: OpenAI’s Codex powered by GPT-5, Anthropic’s Claude Code with Opus 4.5, Google’s Gemini CLI, and Mistral Vibe. Each of these agents was tasked with directly editing HTML and scripting files on a local machine, while a supervising AI model parsed the prompt and delegated coding responsibilities to the various language models, which were equipped to execute the necessary coding actions. Importantly, the development process was kept private, with no insider access granted to the companies involved, ensuring an unbiased evaluation. After the completion of the projects, Ars Senior Gaming Editor Kyle Orland, who is well-versed in Minesweeper, assessed each entry without knowledge of which AI created which version. This evaluation focused on the raw output of the AI models, employing a "single shot" approach to gauge their performance without any human debugging. In practical scenarios, however, any complex AI-generated code typically undergoes some level of scrutiny and refinement by a human programmer to rectify issues and enhance efficiency.