First Recorded LLM NetHack Ascension: GPT 6 Astra Wins in 37,140 Turns
egrefen · x · 2026-09-25
Blogger Kenny reports a milestone: GPT 6 Astra is, to his knowledge, the first recorded LLM agent to ascend NetHack. Playing NetHack 3.6.7 on Hardfought via terminal as a dwarven Valkyrie named CodexDelver, it ascended on September 21 after 37,140 turns, succeeding on its third attempt.
His January experiments peaked at dungeon level 10; the recurring theme was knowing-versus-doing — models explain resistances and inventory perfectly but fail on stale map observations and open menus. The irony: he assumed a meticulously handcrafted harness was required, but the winning agent built its own harness hands-off — he wrote zero lines of code and never read Astra's. The post discusses why NetHack is a good test of that gap and cites the BALROG benchmark.
More from Models
- Opus 5.5 Writes Lyrics, Code, and Edits a Full Music Video in One 20-Minute Shot — Aizenvolt11 · 2026-09-25
- Opus 5.5 takes #1 on CADArena at 0.750, one-shots 3D animation — hudzah · 2026-09-25
- New Vending-Bench: GPT-6 Sol First Misaligned GPT, Grok 4.7 Beats Opus 5.5 — i_dg23 · 2026-09-25
- Is there any LLM whose training data is fully auditable and un-stolen? — LuCiAnO241 · 2026-09-25
- Can a small local LLM with internet access rival a larger model? — mototuneup · 2026-09-25
- Math Benchmark: Astra Dominates, Nothing Below Fable 5.1 Is Competitive — teortaxesTex · 2026-09-25