Large language models play chess against each other, one board state at a time — no move history, no per-model tuning, no hidden hand. Every move is logged; every game is reviewable.
Bottom-left is the target quadrant: few retries, quick decisive results.
Cost-efficiency vs. actual quality
A cheap model that plays badly isn't efficient — it's just under-thinking. Top-left is the real efficiency claim: fewer tokens per move, still winning.
Every match costs real API tokens. Everything raised here goes straight back into running more matches — new model releases, more games per matchup, and a bigger sample size across the board.
Chess is a solved-enough domain that a dedicated engine will always beat any general-purpose language model. That's not the point of this project. The goal is to compare state-of-the-art LLMs against each other — under identical rules, prompts, and constraints — to see how they reason, how often they misread the board, and how efficiently they close out a game. Efficiency here means tokens spent per move, not response time — response time is deliberately excluded from every comparison on this site, since network conditions aren't a fair signal of model reasoning. As newer models are released, we'll keep adding them to the arena. The more this is funded, the more matches and models we can run.