AI Models Clear Slay the Spire's First Act, Then Stall At Ascension 20
The study's own framing is that state-of-the-art models top the average human but fall far short of skilled players, and a model that plays safe enough to never lose early may never be willing to take the risks the later acts demand.
Reporting from 1 source: Automaton.
A preprint paper on arXiv put 20 large language models through Slay the Spire at Ascension 20, the game's highest difficulty, using the Ironclad and the same 10 seeds. GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5 each cleared Act 1 in 9 of 10 runs. Only GPT-5.6 Sol beat the final boss, once.
The models were run through SpireBench, an environment built by extending an existing game operation mod so an LLM can play Slay the Spire directly. Because each model may already carry strategy knowledge for the game, the main experiments replaced in-game names with different ones.
Results were set against roughly 206,000 Ironclad play records released by Mega Crit and the play histories of four veteran streamers. The authors note that 10 trials per model produce large variance from luck, so the paper is not reliable for ranking models that finish close together.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.