Epoch AI launches SpireClimb benchmark as frontier models clear 90% of easy runs

Six months ago frontier models could barely finish a run of Slay the Spire. Now, according to Epoch AI's new SpireClimb benchmark, they win more than 90% of runs on the lowest difficulty setting and reach roughly 20% on the hardest. The jump reflects a sharp improvement in long-horizon planning and resource management under uncertainty, not just pattern recognition.

In SpireClimb the models operate through the same graphical interface a human player uses: they read on-screen text, click cards and choose paths on the map. A few months ago, in that identical setup, models struggled to extract relevant information, missed card clicks and forgot to check critical UI regions. Today the leading models gather strategic information systematically and execute multi-step lines of play without stalling.

Despite the progress, both games remain a long way from peak human performance at the highest difficulty. In the original game GPT-6.1 Sol leads with a 20% win rate on A20 — five times the average player's 4%, based on an analysis of 206 thousand Ironclad runs published by Mega Crit and examined in the SpireBench paper (arXiv:2610.12596) — but well short of top streamers, who sit between 46% and 65%. In the sequel, generally considered easier, GPT-6 Astra leads yet still trails the human elite by a wide margin. One striking detail from the paper: models choose to rest at campfires 57–75% of the time versus 9% for humans, a gap that points to a markedly more conservative approach to health management.

The visual interface may cap performance, but Epoch AI estimates it does not severely handicap the top models, given their 90%-plus success rate on easy. The organization plans to release more than a thousand videos documenting the gameplay captured for the benchmark, a move that will let researchers analyze behavior at far higher resolution than a final score alone. The SpireBench paper remains a pre-print and has not yet undergone peer review, so the figures may shift after formal scrutiny.