Frontier models still trail top players in Slay the Spire

A new benchmark from Epoch AI called SpireClimb shows that frontier language models win more than 90 percent of runs on the easiest difficulty of Slay the Spire, but their win rate on the hardest difficulty stalls around 20 percent. GPT-6.1 Sol leads on the original game, while GPT-6 Astra leads on the sequel, which is considered easier; neither model approaches the performance of top human players.

Six months ago the same models could barely finish a single run. Now, operating in a graphical interface identical to the one a human player uses, they read on-screen text, click cards with precision, and scan relevant UI regions without forgetting critical details. The researchers note that the best models gather strategic information systematically and execute complex sequences smoothly. They plan to release more than a thousand videos documenting the measured games.

The 90 percent win rate on the easiest setting indicates that a visual interface by itself is not a severe limitation. Yet on the highest difficulty the win rate remains far below that of human experts, and the gap is stable across both titles.

Epoch AI's analysis shows that the win rate on hard difficulty decays exponentially as episode length increases. The researchers say this pattern points to search depth — the ability to plan over long horizons and accumulate decisions without small errors cascading into failure — as the bottleneck, rather than the quality of the model's base policy.