Epoch AI said Wednesday that frontier AI models now win more than 90% of Slay the Spire runs at the easiest difficulty, probing multistep planning under uncertainty.
Key Takeaways
Frontier AI models now win more than 90% of Slay the Spire runs at the easiest difficulty, Epoch AI reported Wednesday
Epoch disclosed a 20% win rate at the hardest difficulty but did not name the specific models tested
Slay the Spire requires models to adapt strategies as randomized card draws, enemy encounters, decks and threats change
Epoch documented a six-month improvement from barely beating the game to dominant performance at the lowest difficulty
Epoch AI Research, the nonprofit known for tracking AI capability trends with reproducible benchmarks, published the SpireClimb results on Wednesday. Slay the Spire is a roguelike card game that requires players to build a deck, manage resources and plan several turns ahead against escalating enemy difficulty, a structure that makes it a useful stand-in for multistep planning under uncertainty.
Unlike chess or Go, the game has randomized card draws and enemy encounters, so a model cannot simply memorize optimal openings.
It has to adapt its strategy mid-run as the deck and threats change.
Epoch did not name the specific models tested in the material available, nor did it disclose exact win rates at every difficulty tier beyond the 90% figure at the easiest setting and 20% at the hardest. The six-month jump from barely beating the deck-building game to dominant performance at the lowest difficulty marks one of the faster capability curves Epoch has documented on a single benchmark.
Why Game Benchmarks Still Matter For Serious AI Work
Benchmarks built on games can look like a sideshow next to coding or math evaluations, but they test something static benchmarks cannot.
Slay the Spire requires a model to track a persistent game state across dozens of turns, update plans when a random card changes the viable path and weigh short-term risk against a long-term win condition, the underlying skills needed for an AI agent managing a multistep business workflow or a robotics task that unfolds over time.
Still, the win rate raises a question about what the benchmark captures, rather than providing a direct measure of real-world agent reliability, a model that cannot plan through a card game’s randomness is unlikely to reliably plan through a messy real-world task with similar uncertainty.
Also Read: HCLTech Becomes Anthropic’s Top Claude Partner, Builds New AI Business Unit
The six-month improvement Epoch documents mirrors a pattern seen elsewhere this year, as models have rapidly climbed agentic benchmarks that reward sustained, multistep reasoning even while progress on some static knowledge tests has leveled off.
That divergence suggests labs have specifically targeted planning and tool-use capabilities in post-training, rather than simply scaling up base model size.
From Static Tests To Dynamic Worlds
AI evaluation has shifted over the past two years from one-shot question-and-answer tests toward benchmarks that unfold over time, like Epoch’s own FrontierMath and the agentic benchmarks labs including OpenAI have used to show models running thousands of autonomous tasks.
Game-based evaluations like SpireClimb sit in that same category because they force a model to commit to a strategy, observe the consequences and adjust rather than answer a single isolated prompt.
That dynamic structure is harder to game through memorized training data, since each run’s card sequence differs. Epoch AI has built a reputation for publishing capability trend data that labs themselves do not, including compute-scaling estimates and independent benchmark reruns across major model releases.
What A Harder Difficulty Tier Would Show
The real test for SpireClimb going forward is whether win rates at the hardest difficulty keep climbing at the same pace as the easy tier did.
A jump from near-zero to 90% at the easiest setting in six months does not guarantee the same curve holds at 20% and above, where randomness and compounding mistakes punish weaker planning far more severely.
Epoch’s next model-by-model breakdown, if published, would show whether the gain is broad across labs or concentrated in one or two frontier releases.
Read Next: Elon Musk’s Grok Voice 3.0 Tops Speech Benchmark, Beats Google And OpenAI