How well do current Claude/GPT models play multiplayer Magic the Gathering?

A developer-built harness called Brewhaven ran 80 bot-only multiplayer Magic: The Gathering games with all models set to high reasoning, and after adjusting for deck strength (17.5% win rate for the weakest deck versus 40% for the strongest), Anthropic's Sonnet 5.5 posted a 65% win rate while Haiku 5.5 and GPT-6 Luna each won only 5% of their matches. The harness's creator won roughly 55% of his own matches against the models, against a 25% baseline, and said Sonnet 5.5's diplomacy, dialogue and threat assessment were strong but that a game with three Sonnet 5.5 opponents costs roughly $8, making high-quality AI Magic play not economically viable at present.

How well do today's frontier language models play card games, and in particular multiplayer Magic: The Gathering? I built Brewhaven, a harness that lets AI agents play card games against each other and against humans. Brewhaven also tests more than game mechanics. Agents have to plan, scheme, bluff, anticipate outcomes, negotiate, persuade, and judge who the real threat is. Unfortunately, a full benchmark where different AI models compete against one another would blow a giant hole in my wallet. I ran a much more limited version, 80 bot-only games with all models set to high reasoning. After adjusting for deck strength weakest deck had 17.5% win rate against 40% for the strongest , the models clearly separated by class, where even Haiku 5.5 and GPT-6 Luna could only win 5% of their matches. The strongest model I ran Sonnet 5.5, had a 65% win rate. However, against me, I won roughly 55% of my matches 25% is the baseline so they're not superhuman yet. Sonnet 5.5's diplomacy and dialogue is stellar, sounding very natural and engaging. It's also fast and makes good threat assessment. I didn't realize how poor threat assessment would annoy me and degrade the game experience, but basically I could feel the difference in the model just by their actions. Downsides is that it costs roughly $8/game with 3 Sonnet 5.5 opponents, so not economically viable at the moment to play a high quality MTG game against AI. I ran out of Claude API credits before I could film the video, so here's me playing against Luna .