Epoch AI's SpireClimb benchmark has AI models play Slay the Spire through computer use. This market is about the Slay the Spire 2, Ascension 10 variant, scored as the win rate across runs.
Current state (Epoch data as of Oct 29, 2026): the top score is GPT-6 Astra (max) at 34.7%, followed by GPT-6.1 Sol (max) at 29.3%. Only 7 models have been scored, and all of those runs were in late September. Several newer models (e.g. GPT-6.1 Terra, Gemini 4 Argon) have not been scored on this variant yet.
Resolution: the highest mean_score in Epoch AI's spireclimb_sts2_a10.csv, from the benchmark data download (https://epoch.ai/data/benchmark_data.zip), or the matching table on Epoch's benchmarking hub if the file moves. I'll check it on or after November 30, 2026, 23:59 UTC. Any model counts, from any lab, as long as Epoch has published the score by then.
Buckets use the raw score (34.7% = 0.3467):
Below 35%: no model has beaten GPT-6 Astra's 0.3467 by more than a rounding hair.
35% to under 45%
45% to under 60%
60% or higher
If Epoch retires the StS2 Ascension 10 variant, I'll use the last published value. If Epoch changes the scoring method in a way that rescores old models, I'll use the new numbers as published on the resolution date. I won't trade in this market.
Background: https://epoch.ai/publications/ai-systems-are-improving-rapidly-at-slay-the-spire-and-its-sequel
1,000My take: about 60% that nothing beats 34.7% by Nov 30, then roughly 22% for 35–45%, 12% for 45–60% and 6% for 60%+.
The reason is cadence, not ability. In Epoch's data file, all seven scored runs on this variant started between Sept 22 and Sept 30. Since then GPT-6.1 Terra has been scored on FrontierMath and Gemini 4 Argon on other Epoch tests, but neither on SpireClimb. So a new record needs two things in about a month: Epoch running the game again, and a model that beats Astra. Astra's 34.7% is also a noisy number (Epoch's standard error is about ±5.5 percentage points), so a new model could clear it by luck.
What would change my mind: Epoch posting a fresh SpireClimb batch, or a big frontier launch in November. Data: https://epoch.ai/data/benchmark_data.zip · write-up: https://epoch.ai/publications/ai-systems-are-improving-rapidly-at-slay-the-spire-and-its-sequel