AI models can now beat Slay the Spire, but only the easy mode

Epoch AI built a benchmark called SpireClimb that has frontier models play the card game Slay the Spire and its sequel through a visual computer-use interface, so the model reads the screen the way a person would. Each model gets 15 single-shot attempts per character, can keep notes across context compaction, and can check the game wiki. At base difficulty the top models now win almost every run: GPT-6 Astra 98%, GPT-6.1 Sol 97%, Claude Opus 5.5 93%. Six months ago they struggled with the same setting.

The gap opens up at hard settings. At the maximum Ascension 20 Heart run for the first game, GPT-6.1 Sol wins 20%, GPT-6 Astra 17%, and every other model lands at 2% or zero, against an estimated 85% for top human players. The sequel, which is newer and mostly absent from training data, tells a cleaner story because the models cannot be leaning on memorized strategy: GPT-6 Astra reaches 35% at Ascension 10 while humans sit near 95%. Epoch notes that strong first-game scores may partly reflect familiarity, and that the vision harness can understate reasoning because interface errors creep in.

The authors are candid about the benchmark's shelf life. If frontier labs start training on these games, the scores could saturate within a few model generations, which would end its usefulness as a measure.

Why it matters

If you are judging how close agents are to doing real multi-step work through a normal screen, this is a concrete gauge: models clear the base game but still fall far short of expert humans on the hard runs. The six-month jump at base difficulty shows the direction of travel, so treat today's ceiling as a moving one when you plan around it.