Epoch AI's SpireClimb benchmark has AI models play both Slay the Spire games through a computer-use harness (they see the screen and click), at the hardest settings: Slay the Spire 1 on Ascension 20 with the Heart kill (60 fixed-seed runs), and Slay the Spire 2 on Ascension 10 (75 fixed-seed runs). The score is the average of the model's win rates in the two games. Leaderboard: https://epoch.ai/benchmarks/spireclimb
State on 2026-11-11 (spireclimb.csv in Epoch's downloadable data, https://epoch.ai/data/benchmark_data.zip):
Top score: GPT-6 Astra (max) at 34.0%, set in early September and unbeaten since.
Next: GPT-6.1 Sol (max) 23.8%, GPT-6.1 Terra (max) 22.3%, Claude Fable 5.2 (max) 19.6% (added Nov 9), GPT-6 Sol (max) 9.2%, Claude Opus 5.5 (max) 5.8%.
Question: what is the highest SpireClimb score any model has at close?
Resolution: On or shortly after Dec 11, 2026, I'll take the maximum score for any model in spireclimb.csv (the mean_score column × 100) from Epoch's benchmark_data.zip, or from the leaderboard page if the download is unavailable. Only runs whose evaluation started on or before Dec 10, 2026 23:59 UTC (the CSV's "Started at" column) count. If the page and the CSV disagree, the CSV wins. A score exactly on a bucket edge goes in the higher bucket (35.0% counts as "35% to under 45%"). If Epoch renames the benchmark, I'll use its direct successor. If Epoch changes the game settings or run count, I'll go by the setup described above where it can be told apart, and explain in the comments before resolving if it can't.
I won't trade in this market.
1,000My take: ~70% GPT-6 Astra's 34% is still on top Dec 10, 15% for 35-45, 10% for 45-60, 5% for 60+. Three newer models have all come in under 34%. The likeliest way up is a new Claude that Epoch runs in time (Fable went from 2% to 20% in one version). Data: https://epoch.ai/data/benchmark_data.zip