Player not loading? Watch on YouTube
This Russian-language comparison is summarized in English from its complete English caption track.
The speaker compares Aider, OpenCode and Pi with the same local LLM, identified in the metadata as Qwen 3.8 27B. Each receives a Markdown specification for a browser game inspired by Crimsonland, where a character shoots approaching monsters and survives as long as possible. Game balance matters alongside working movement and weapons.
This is an informal coding assistant comparison. The speaker explains that follow-up prompts differed because each agent produced different bugs; he instead spent roughly similar amounts of time on them. He reports spending about 2 million tokens with Aider without getting a working prototype after roughly an hour, while acknowledging that his configuration or familiarity may explain the result.
OpenCode produced a playable first result with minor interface bugs, followed by several dozen refinement iterations and about 6 million tokens overall. The speaker describes its terminal interface, built-in LSP support and MCP and plugin options. Pi starts with neither LSP nor connected MCPs, but supports extensions. At about 3.2 million tokens, it produced the game he preferred for its balance and gameplay.
A separate model-generated code review also favored Pi. These judgments reflect this experiment rather than a general ranking. The speaker notes that the specification did not request tests and the resulting code had none.
