Six AI models played chess through CLIs without checkmate

Building a self-hosted workspace for running AI coding CLIs side by side recently inspired a unique experiment. On October 18, the developer behind SpaceApp arranged for six artificial intelligence models to play chess against each other, with each model making its moves exclusively through its own command-line interface. Surprisingly, across a total of twelve games, not a single model succeeded in delivering a checkmate. Instead, the most fascinating takeaway lies in the specific ways they lost.

While twelve games represent a relatively small sample size and a tournament decided mostly by forfeits reveals very little about traditional chess proficiency, it highlights a different critical metric. The matches demonstrated how reliably a model can respond through its CLI when every single turn is bound by strict deadlines.

This reliability ensures zero timeouts, zero provider errors, and zero illegal moves, game after game. Such consistency is precisely the property developers care about most when AI agents are left to run unattended for hours, differing substantially from rankings derived from single-prompt leaderboards.

Stock photo for illustration only, not from the actual event