Slay the Spire AI win rates: Epoch's SpireClimb shows 90%-plus at base, about 20% at max difficulty

Frontier AI models now win Slay the Spire at its base difficulty at least 90% of the time, and the best of them reach about one win in five at the game's hardest setting, according to Epoch AI's SpireClimb benchmark, published October 14, 2026. Epoch says six months earlier models struggled to win at all. Top human players still sit far ahead on the hardest setting, and Epoch itself warns the test may saturate within a few model generations.

What SpireClimb measures

Slay the Spire is a card-based roguelike: you build a deck, fight through three acts and beat a final boss. A single mistake in building the deck can sink a run hours later, which makes it a long-horizon planning test with hidden information and heavy randomness. Epoch's report, by Benjamin Ou and Simon Chen, covers the original game and its sequel, Slay the Spire 2, with models playing through a visual computer-use harness, meaning they see the screen and click as a person would.

The harness choice cuts both ways. Epoch says visual computer use may hide some of a model's reasoning ability, since reading a screen is its own source of errors. So scores are a floor on what the model could do with a cleaner interface, not a ceiling.

The results

At the base difficulty of the original game, known as Ascension 0, Epoch reports frontier models win at least 90% of the time. The top scores are GPT-6 Astra at 98% and GPT-6.1 Sol at 97%.

The hardest setting, Ascension 20 with the Heart as the final boss, is a different test. Only two models won more than one run: GPT-6.1 Sol at 20% and GPT-6 Astra at 17%. Epoch estimates top human players at about 85% there.

For Slay the Spire 2 at Ascension 10, models did better than at a comparable level of the original. GPT-6 Astra reached 35% and GPT-6.1 Sol 29%, against an estimated 95% for top humans. The authors note the sequel is generally considered easier, which is part of why the numbers look better.

Read the small numbers carefully

Secondary coverage of the report adds detail we could not check against Epoch's own tables. AlphaSignal reports that Claude Opus 5.5 also scored at least 90% at base difficulty but 0% at Ascension 20 with the Heart. It also reports that GPT-6.1 Sol won 33% of runs as the Watcher and 27% as the Ironclad but none in 15 attempts as the Silent, and that Epoch has no clear explanation for the Silent result.

It also states the sample size: each model gets 15 attempts per character, so 60 attempts in the original game. One extra win moves the rate by about 1.7 points. That means the gap between 17% and 20% at the hardest setting is a few runs, not a ranking. The right reading is that two models cleared a very hard bar sometimes and the rest did not.

On the human side, AlphaSignal says strong players regularly report Ascension 20 Heart win rates above 50%, with one player reporting 80% to 95%, but notes these come from broader play, not Epoch's fixed seeds. Epoch's own estimate is about 85%. The point either way is that humans remain well ahead.

Why a game matters

Games are a legible way to test the skills agents need in the real world: planning across many steps, managing resources, recovering from early mistakes, and staying coherent for hours. The jump from struggling to win at all to 90% or better at base difficulty in six months is the kind of curve that our explainer on benchmark saturation warns about from the other side: once a test is solved, it stops separating models.

Epoch says the authors expect the benchmark could saturate within a few model generations, especially if labs train on these games. That caution is the most useful line in the report. A game with a fixed ruleset and public guides is easy to train on directly, and a score that comes from training on the test says less about general planning.

It also differs from the benchmarks that grab headlines. We covered the gap between easy-looking tasks and hard ones in our robot tasks piece, where frontier models solved 21 of 84 tasks. Slay the Spire is on the favorable end of that spread: a closed world, clear rules, and a screen to read.

What this does and does not show

It shows steady progress on long-horizon play through a visual interface, and that the best OpenAI models lead this particular test. It does not show that a model can run a business, manage a project or plan in an open world. It also does not settle a model ranking: the two top models are within three points at base difficulty, which is within the noise a 60-run sample allows. For how to read any single score, see our guide to what AI leaderboards rank.

Our take

Treat SpireClimb as a good progress gauge and a poor ranking tool. The 90% mark at base difficulty is real and recent. The hard-mode gap to humans, about 20% for the best model against an estimated 85% for top players, is the figure to track, along with whether it closes as fast as the easy-mode gap did. Watch for Epoch's next run and for any lab publishing its own SpireClimb training. Epoch's report contains no quote we can verify word for word, so none is used here.