Epoch AI's SpireClimb: models beat human author at Slay the Spire but still err each act
Jsevillamol · x · 2026-10-17
Epoch AI recorded over a thousand runs of Slay the Spire and its sequel for its new benchmark SpireClimb, publishing reconstructed videos with the models' in-game thinking as subtitles. Reviewer Jsevillamol notes models now achieve better win rates than he does, but he still spots clear mistakes about once per act, and considers their deck-building choices subpar. Epoch won't update the playlist with future models but hopes the archive offers insight into how current models play.
More from Models
- Stanford professor: Astra is a poor writer, Gemini 3 Pro was the high-water mark — anshulkundaje · 2026-10-17
- Researchers slam frontier LLM writing quality: "getting worse with every version" — arjunrajlab · 2026-10-17
- Perplexity quietly retires GLM 5.3, user still used it two minutes earlier — ChrisUniverse · 2026-10-17
- NVIDIA Launches Fastino Training API: Pay Per Token for SFT and RL — NVIDIAAI · 2026-10-17
- F5 Post-Trains Nemotron for Enterprises: Frontier Accuracy at 40x Lower Cost — NVIDIAAI · 2026-10-17
- Post-Trained Nemotron Writes CUDA Kernels: KernelBench L2 22.8% to 82.1% — NVIDIAAI · 2026-10-17