Astra tops VoxelBench 3D leaderboard

Ananth @Ananth7eOriginally posted 2026.09.05

Open original on X

so much for "gpt models are bad at 3d generation and spatial reasoning" gpt-6 astra ranked #1 on voxelbench with the highest score the benchmark has ever seen. >2620 elo >3.5k+ votes > 328 elo ahead of opus for context, the best model on the leaderboard back in april, gpt-5.5 pro was sitting around 2200 elo, and opus 5 was at 2292. so that's ~400 elo jump from sol

This post quotes another post: x.com/voxelbench/status/2096186872384942345