Epoch AI's SpireClimb benchmark tests AI models on Slay the Spire 1 and 2
Article timeline
The articles in this story. Times are ET.
Epoch AI@EpochAIResearchOfficialEpoch AI says visual interface may limit frontier model game performanceAIEpoch AI says the visual interface could be limiting model performance in its game benchmark. Frontier models still win over 90% of games on the easiest difficulty, so the interface does not appear to severely hamper them. Epoch AI says measuring how models play under human-like conditions is worthwhile regardless.
- Epoch AI@EpochAIResearchOfficialEpoch AI says models now play SpireClimb through a graphical interface like humans
AIEpoch AI says models now play SpireClimb through a graphical interface, the same way humans do. A few months ago, models in a similar setup failed to read on-screen text, clicked cards inaccurately, and forgot to check key interface elements.
- Epoch AI@EpochAIResearchOfficialFrontier models score higher on game sequel but trail top humans
AIEpoch AI reports that frontier model win rates on the sequel, which is broadly considered easier, are somewhat higher than on the first game. The leading model, GPT-6 Astra, still performs significantly worse than top human players.
- Epoch AI@EpochAIResearchOfficialFrontier models still trail top human players on hardest difficulty
AIEpoch AI reports that frontier models are improving quickly, but their win rates on the hardest difficulty in both games remain far below those of top human players. GPT-6.1 Sol leads in the original game.
- Epoch AI@EpochAIResearchOfficialEpoch AI's SpireClimb shows frontier models now win 20% of hardest Slay the Spire runs
AIEpoch AI says its new SpireClimb benchmark shows frontier AI models now win over 90% of Slay the Spire runs at the easiest difficulty. At the hardest difficulty, they win up to 20% of runs. Six months ago, the same models could barely beat the game.
- Epoch AIOfficialFrontier AI models now win Slay the Spire at high rates on the SpireClimb benchmark
AIEpoch AI's SpireClimb benchmark finds frontier models beat Slay the Spire 1 at base difficulty (Ascension 0) at least 90% of the time, up from struggling six months earlier. GPT-6 Astra reached a 17% win rate and GPT-6.1 Sol a 20% win rate at maximum difficulty (Ascension 20 Heart), far below top human players at 80% or more. Models also won more often on Slay the Spire 2 at Ascension 10 than on the comparable original tier, though the authors note that game is easier.