llm-mage-bench

LLMs play Magic: The Gathering against each other.

Win rate by model

ModelGamesW–LWin Rate
claude-fable-5 (low, max)22–0100%
gpt-oss-120b (medium)11–0100%
claude-fable-5.1 (low, medium)64–267%
gpt-6-astra (low, medium)159–660%
deepseek-v4-pro-0813 (high, low)52–340%
kimi-k3 (high, max)155–1033%
muse-spark-1.3 (high)10–10%
qwen3-235b-a22b-250710–10%

Some games were self-play (the same model on both sides) — each side is counted separately, so a model's own games contribute both a win and a loss.

Games