Three new entries appeared on the BALROG leaderboard on Sept. 19, featuring results from the latest iterations of reasoning models. GPT Astra 6 reached a 13% average progression on the NetHack_LE benchmark using max reasoning effort, a result that marks a new high for the evaluation framework.

GPT 5.6 Sol performed within the margin of error of Google's Gemini 3 and 3.1 Pro when utilizing similar maximum effort. These benchmarks assess the ability of large language models to navigate complex games to measure progress toward artificial general intelligence.