A study not yet peer-reviewed challenged 20 large language models with Slay the Spire's highest difficulty, Ascension 20. They attempted a total of 200 runs, with only one run by GPT-5.6 Sol defeating the final boss.
The study uniformly used the Ironclad character, having each model attempt 10 runs under the same 10 random seeds. The experiments were conducted through an environment called SpireBench; the main experiments also replaced in-game names to account for the possibility that models had learned Slay the Spire guides.
Some models performed strongly in Act 1: GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5 all cleared Act 1 in 9 out of 10 runs, but the number of runs defeating the final boss dropped significantly in later acts. The study also compared model performance against approximately 206,000 recorded Ironclad runs and gameplay records from four veteran streamers. The paper suggests that while stronger models can handle immediate combat, calculate damage, and manage energy, they still lag significantly behind experienced players in making choices that build strength for subsequent acts.
The study observed that models tended to avoid elite encounters, heal at rest sites, and were less likely to skip post-combat card rewards. Calculated by runs reaching the Act 1 boss, models averaged 1.25 elite encounters, compared to 2.6 for veteran streamers; models chose to rest 68% of the time, versus 9% for veteran streamers. Models arrived at the Act 1 boss with a higher average proportion of health remaining, but this may have caused them to miss opportunities for powerful cards or other rewards.
The paper states that repeated playthroughs did not lead to clear performance improvements. However, each model only attempted 10 runs, and random fluctuations may have affected the rankings of models with similar scores; the random seeds used by humans and models were also different, and the experiment only tested the Ironclad character. These results cannot yet be used to strictly rank the strength of models versus human players.