Epoch AI launches SpireClimb: frontier models win 90% of base Slay the Spire runs, 20% at max

What Epoch AI published

Epoch AI, an independent research nonprofit, published the first results from its SpireClimb benchmark on 14 October 2026. The report by Benjamin Ou and Simon Chen tests how well AI models play Slay the Spire, a single-player deck-building game from 2017, and its early-access sequel, Slay the Spire 2. Epoch says the games have long resisted both classic machine learning and large language models because of randomness, hidden information and a huge decision tree.

Epoch's headline finding is speed of progress. Frontier models now beat the first game at its base difficulty, Ascension 0, at least 90% of the time. Six months earlier, Epoch says, they struggled to win at that level at all.

How SpireClimb works

Models play through a visual computer-use harness based on Inspect ReAct, so they see and click the game interface as a human would. Epoch says the aim is to measure what an unskilled user would get by opening a desktop AI app and asking it to play the game. Each model gets 15 independent single-shot attempts per character in each game, on a fixed set of seeds. Models can take notes that survive context compaction and can consult the game's wiki.

Because base difficulty is already mostly solved, SpireClimb targets the hardest settings: Ascension 20 with the optional final boss, The Heart, in the first game, and Ascension 10 in the sequel. Epoch froze Slay the Spire 2 at version v0.107.1 so that results stay comparable.

The results by model

On the first game's hardest setting, GPT-6.1 Sol won 20% of runs and GPT-6 Astra won 17%. Epoch says this is up from 0% a few months ago. Sol did better with the Ironclad and Watcher characters than with Silent and Defect, which roughly mirrors the record of XecnaR, a player Epoch describes as widely considered the best. Top players regularly win more than 50% at this level, and XecnaR wins 80% to 95% depending on the character.

On Slay the Spire 2, the top models passed 25% at Ascension 10. Epoch notes that the sequel is easier at maximum difficulty, since it has no equivalent of The Heart yet, and that hundreds of human players sustain win rates above 80% there. Because the sequel is new enough to be largely outside training data, Epoch reads the results as evidence that frontier models do not lose much when they play a less familiar game.

Limits and what to watch next

Epoch says the visual harness probably understates pure reasoning skill, because interface mistakes still occur. It also warns that labs may train on Slay the Spire, so the benchmark could saturate within a few model generations. Epoch plans to test new models as they appear and to add more games. Other benchmark results are listed under AI Models.

Questions people ask

What is SpireClimb?

A benchmark from Epoch AI in which AI models play Slay the Spire and Slay the Spire 2 at maximum difficulty through a visual computer-use harness, with 15 attempts per character.

Which model scored highest on SpireClimb?

On the first game's hardest setting, GPT-6.1 Sol won 20% of runs and GPT-6 Astra 17%, according to Epoch AI's 14 October 2026 report.

How do AI models compare with human players?

Epoch AI says top human players win at least 80% of games at maximum difficulty in both games, far above the best model results.

Sources

1. AI systems are improving rapidly at Slay the Spire and its sequel (Epoch AI, 14 Oct 2026)