Researchers built a benchmark that makes AI models play an entire game of Slay the Spire, and most of them choke by the end.
The benchmark, called SpireBench, tested 20 large language models on the hardest difficulty of the commercial deck-building game, where early choices shape whether a run survives dozens of fights later. Frontier models beat simple heuristic baselines and matched the average human player, but stayed well behind expert players. The gap grew as runs went on: top models won the first act about as often as experts, then fell off. The researchers also found the models misjudge risk throughout a run, hoarding health instead of spending it to get stronger, and a test that renamed every game term showed some of their performance comes from memorized knowledge of Slay the Spire rather than in-game reasoning.
This matters because most agent benchmarks are short tasks that end before an AI's mistakes compound. Slay the Spire forces decisions to pay off, or not, over a full run, which is closer to how real strategic work unfolds. The researchers also found models don't get better by using notes they write for themselves between runs, since those notes are often wrong and overly cautious, which pokes a hole in the idea that agents can simply accumulate useful memory over time.
Slay the Spire has become an odd but useful proving ground for AI planning research precisely because its difficulty comes from long-term tradeoffs, not twitch reflexes. An AI that can win the first act but can't close out a run looks a lot like a chess engine that plays a strong opening and then blunders the endgame.