SpireClimb
Games · 2026-10-14
Epoch AI's SpireClimb has models play Slay the Spire 1 and Slay the Spire 2 at maximum difficulty through a screenshot-based computer-use harness built on Inspect ReAct, at each model's maximum reasoning setting. Slay the Spire 1 is played at Ascension 20 with the Heart kill (four characters, 15 fixed seeds each, 60 runs) and Slay the Spire 2 v0.107.1 at Ascension 10 (five characters, 15 seeds each, 75 runs). Each run is a single attempt. The score is the equal-weight mean of the two games' win rates. Top human players win about 85% and 95% of these runs.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 25.7 |
| GPT-6.1 Sol | 24.7 |
| GPT-6 Sol | 7.5 |
| Claude Opus 5.5 | 3.3 |
| Claude Fable 5.1 | 2.0 |
| Claude Sonnet 5.5 | 1.3 |
| GPT-6 Luna | 0.0 |