Epoch AI's SpireClimb shows frontier models now win Slay the Spire
- Source
- x.com
- Date
Epoch AI's SpireClimb benchmark has frontier models play the deckbuilding game Slay the Spire using only its graphical interface, with no special API into game state. That makes it a long-horizon computer use test: a run requires many sequential decisions where early choices shape later options. Epoch reports win rates above 90% on the easy setting and about 20% on the hardest difficulty. Epoch describes this as a large jump from results six months earlier, while noting that the hardest setting remains well below top human players. The gap between easy and hard difficulty is the useful part for agent builders. Models now handle long GUI sessions reliably when the margin for error is wide, but still fall off when sustained planning under tight constraints decides the outcome.
Six months ago, frontier AI models could barely beat Slay the Spire. Now, according to our new SpireClimb benchmark, they win 90%+ of runs at the easiest difficulty, and up to 20% of runs at the hardest.
Context
SpireClimb shows a large gap between beating Slay the Spire's easiest setting and mastering its hardest one. The 's launch report records win rates above 90% for several leading models at base difficulty, but a best result of 20% at maximum difficulty with the optional final boss. Top human players remain substantially ahead at that setting. Launch analysis.
The models play through screenshots and mouse-and-keyboard actions, with access to the game's wiki. Each model faces the same random seeds. This tests whether it can turn available information into successful decisions across a whole run, while controlling some differences in game knowledge and luck. Benchmark methodology.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Checking sign-in…
Loading comments…