GPT-6 Astra Ascends NetHack on Third Attempt in First Recorded LLM Run

Blogger Graham Lee (gleech, with Kenny documenting the run) announced a milestone: an agent codenamed CodexDelver, powered by GPT-6 Astra, achieved a NetHack ascension — the first recorded LLM agent to beat the game. Ethan Mollick noted the model succeeded on only its third attempt, calling it a stunning achievement, though questions about training-data contamination followed.

Confirmed

The agent CodexDelver, driven by GPT-6 Astra via a remote terminal, played NetHack 3.6.7 on the public Hardfought server and completed an ascension; blogger Kenny documented the full run.
The character was a dwarven valkyrie; gleech announced completion in September 2026, with the run starting September 21 and made public around September 25; teortaxesTex relayed that the team (Kenneth Bergquist) claimed the first AI ascension.
gleech's announced stats: 176,000 turns and 1.76 million points.
Mollick relayed this as the first recorded LLM completion of NetHack, adding that he himself has played for years without ever ascending.
teortaxesTex (repost) added context: NetHack has long been treated as a jokingly "AGI-complete" benchmark that even humans find extremely hard; gleech's team's earlier tests of Astra had all failed.
The blogger says he also wrote the agent code himself.

Unconfirmed

Figures conflict: one summary cites 37,140 turns, another 176,000 turns and 1.76 million points; exact stats await confirmation from the original blog.
Researcher pfau raised a key transparency question about whether the training corpus contained NetHack completion records (contamination), which remains unresolved.

Why it matters

NetHack is one of the hardest classic roguelikes and a semi-serious yardstick for general intelligence; the first LLM ascension is a landmark case of long-horizon planning and persistence. The contamination question, however, determines whether the feat reflects generalization or memorization.

2026-09-25 ~ 2026-09-27 · 7 related posts

Primary sources

First recorded LLM NetHack ascension: GPT 6 Astra scores 1.77M points over 37,140 turns — gleech · 2026-09-26 06:50 (UTC+08:00)
First Recorded LLM NetHack Ascension: GPT 6 Astra Wins in 37,140 Turns — egrefen · 2026-09-25 08:03 (UTC+08:00)
GPT-6 Astra beats NetHack on third try, first recorded LLM agent ascension — emollick · 2026-09-25 10:18 (UTC+08:00)

[source] First Recorded LLM NetHack Ascension: GPT 6 Astra Wins in 37,140 Turns — egrefen · 2026-09-25
Ethan Mollick: GPT-6 Astra beats Nethack on just its 3rd try — emollick · 2026-09-25
[source] First recorded LLM NetHack ascension: GPT 6 Astra scores 1.77M points over 37,140 turns — gleech · 2026-09-26
First documented AI ascension run claims the Nethack 'AGI-complete' crown — teortaxesTex · 2026-09-26
GPT-6 Astra beats Nethack on third try, sparking training-data transparency concerns — PMinervini · 2026-09-27

2 near-duplicate retellings: emollick · gleech