WHAT NEW IN X11.02 MON

AI collections from X · 22 collections · 64,204 posts

gpt-6THREAD

GPT-6 Astra nearly doubles score with code execution on MazeBench

A comparison chart shows GPT-6 Astra with code execution nearly doubling its score, and Astra without code still beating GPT-5.6 Sol with code.

💺
@patience_caveArchived 2026.09.15Originally posted 2026.09.14
THREADPosted in 2 parts
  1. GPT-6 Astra with code execution saw a massive leap over GPT-5.6 Sol with code execution, nearly doubling its score. A big surprise though was GPT-6 Astra without code surpassed GPT-5.6 Sol with code. Astra is a powerful agent.

    GPT-6 Astra nearly doubles score with code execution on MazeBench (thread part 1)
  2. MazeBench vs GPT-6 Astra + Python After 400 million tokens and 24 hours of reasoning, the final score was 23% Without Python, Astra scores only 14% in this 3D environment

    GPT-6 Astra nearly doubles score with code execution on MazeBench (thread part 2)
💺 @patience_caveOriginally posted 2026.09.14
Open original on X
x.com/patience_cave/status/2099518168029106196Archived 2026.09.15 10:10