Study: 20 LLMs Play Slay the Spire at Ascension 20, One Boss Kill
A preprint on arXiv had 20 large language models play Slay the Spire at Ascension 20; they cleared Act 1 but struggled later, with one final boss defeat.
A preprint research paper published on arXiv describes a study in which 20 large language models played Slay the Spire at Ascension 20, the game's highest difficulty. Each model played 10 runs using the same 10 seeds, all as the character Ironclad.
The experiment used a custom environment called SpireBench, built by extending an existing game-control mod so the models could play Slay the Spire directly. In the main experiment, in-game names were replaced to account for the possibility that the models had learned Slay the Spire strategy information.
GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5 each cleared Act 1 in 90% of runs. Act 2 clear rates were 50% for GPT-6 Astra and Claude Fable 5.1 and 40% for Claude Opus 5. Act 3 clear rates were 20% for Claude Fable 5.1 and 10% each for GPT-6 Astra and GPT-5.6 Sol. GPT-5.6 Sol defeated the final boss in 10% of runs; across all 20 models and 10 runs each, the final boss was defeated only once.
The paper compared the results with about 206,000 Ironclad play records published by Mega Crit and with the play histories of four veteran Slay the Spire streamers. Among the Mega Crit records, the Act 1 clear rate was 47%, Act 2 was 15%, Act 3 was 7% and final boss defeat was 4%. For the four veteran streamers, Act 1 was around 90%, Act 2 around 70%, Act 3 56-71% and final boss defeat 46-65%.
The paper suggests top models handle short-term combat well, using energy efficiently and finishing fights in fewer turns, but were much worse at long-term strategy than veteran streamers, including route selection and event choices. Models tended to prioritize protecting HP and avoided risky choices that could lead to future growth. For example, at rest sites, rest was chosen 68% of the time by the models, 9% by veteran streamers and 23% by all humans.
Slay the Spire began early access on Steam in 2017 and was formally released in 2019; a sequel, Slay the Spire 2, began early access on Steam in March 2026. The study's limitations include no confirmation of peer review, only 10 trials per model, humans and models not playing the same random seeds, and only the Ironclad being compared.
Quick answers
What game did the LLMs play in the study?
They played Slay the Spire at Ascension 20, the highest difficulty, all as the character Ironclad.
How did the LLMs perform compared with human players?
They cleared Act 1 easily, but their later-act clear rates and final boss defeat rate were below veteran streamers and, for some metrics, below broader human records.
Was the paper peer-reviewed?
The paper is a preprint on arXiv with no confirmation of peer review, and it lists other limitations such as only 10 trials per model.