On October 14, 2026, Epoch AI published “AI systems are improving rapidly at Slay the Spire and its sequel” by Benjamin Ou and Simon Chen, introducing a new benchmark, SpireClimb. Slay the Spire is a single-player deck-building game with randomness, hidden information and a huge decision tree, long resistant to both classic machine learning and language models. In June 2026 Anthropic described Claude Fable 5 as reaching the game’s final act three times more often than its predecessor, wording that implied it still could not reliably win.
SpireClimb plays both Slay the Spire 1 and its recent sequel through a single-agent visual computer-use harness, approximating what an unskilled user would get by telling a desktop AI app to play. Each model gets 15 independent single-shot attempts per character on fixed seeds, may keep notes across context compaction, and may consult the community wiki. Epoch reports that frontier models now win the first game at base difficulty (Ascension 0) at least 90 percent of the time, where six months ago they struggled to win at all. At the maximum setting, Ascension 20 with the optional final boss, GPT-6 Astra won 17 percent and GPT-6.1 Sol 20 percent, up from 0 a few months earlier. On the sequel’s hardest setting (Ascension 10) top models exceeded 25 percent.
Because the sequel is new enough to be largely absent from training data, Epoch uses it as a check on whether skill generalizes beyond memorized strategy for the decade-old original, about which every model tested showed extensive prior knowledge.
Why it matters: it is a cheap, legible, long-horizon test that combines planning under uncertainty with real computer use, and the jump from 0 to 20 percent at maximum difficulty in months is a sharp capability signal.
What it does not show: top human players sustain 80 percent or more at maximum difficulty in both games, so models remain far behind experts. Epoch cannot rule out that labs train on these games, samples per character are small, and the sequel is in early access, so Epoch froze it at one version.