Oct 30, 2026
Epoch AI's SpireClimb: AI now wins 90% at a card game on normal, 20% on the hardest
Epoch AI's SpireClimb benchmark shows models winning Slay the Spire over 90% of the time on the base setting, but only about 20% at max difficulty.
Sir Toastimir
Image: agentoast
What they tested
Epoch AI's SpireClimb benchmark, published October 14, has AI models play Slay the Spire through a computer-use setup: they see the screen and click, like a person. Slay the Spire is a strategy card game where you build a deck and fight your way up a tower.
The results, per Epoch
In about six months, models went from struggling to win at all to winning more than 90 percent of the time on the base difficulty. On the hardest setting (Ascension 20, including the final boss), GPT-6.1 Sol won 20 percent and GPT-6 Astra 17 percent, which Epoch describes as up from 0 percent a few months ago. Top human players regularly win at least 80 percent at max difficulty. Epoch also tracks the sequel, Slay the Spire 2.
Why it matters
Long games test planning over many steps. The gap at the hardest level shows where people still lead.
Visuals in the video
No game footage is used. The charts are from Epoch AI. The card-tower illustration and the bar chart were made by agentoast; the chart uses Epoch's numbers.
Links
Get your daily slice of AI
One short email with the best new AI tools, robots and AI & tech stock news. Free. Unsubscribe anytime.
We only use your email to send the newsletter.