GPT-6 'Astra' Does 34 Math Steps in Latent Space, 4x More Than Sol
MaartenBaert · reddit · 2026-09-10
Developer Maarten Baert built LatentMathBench to test latent-space reasoning without chain-of-thought, after rumors that GPT-6 'Astra' uses recurrent depth. Astra completed 34 consecutive simple math operations without CoT — 4x more than Sol (8) and far ahead of Claude Opus 4.6 (12) and 24 other models. The benchmark had seen little progress until Astra's release, suggesting a real architectural jump.
Related event: GPT-6 Astra does 34-step mental math without chain-of-thought(2 posts)→
More from Models
- Claude usage limit glitch fixed, but reset dates pushed from Sep 13 to Sep 15 — ___Patrice___ · 2026-09-10
- Astra beats Celeste chapter 1 from screen input and mines a Minecraft diamond overnight — imjustnewatai · 2026-09-10
- Letting GPT-6 Astra drive software UIs directly beats built-in agents, dev reports — brandon_galang · 2026-09-10
- Mathematician alleges OpenAI trained Astra on his unpublished work on Gromov's soficity conjecture — ValerioCapraro · 2026-09-10
- Users joke Claude is pushing them toward femboy personas — shakoistsLog · 2026-09-10
- Astra does computer use with zero GUI input by hooking internal functions directly — pokeuser61 · 2026-09-10