GPT-6-Astra reportedly can't run without CoT; even 'low' effort hits 99.6% multiplication accuracy
maksym_andr · x · 2026-09-17
In a follow-up to his multiplication eval, maksymandr found GPT-6-Astra cannot run without CoT — the lowest available effort setting is 'low', yet it still achieves near-perfect 99.6% multiplication accuracy at that setting.
This complements his earlier finding that Claude Opus 5 hits 100% on 20x20-digit multiplication at max reasoning, framing no-CoT capability as a key metric for recurrent-depth architectures.
Related event: Claude Opus 5 Aces 20×20-Digit Multiplication, 1200/1200 Without Tools(7 posts)→
More from Models
- Filtering synthetic envs where Qwen deterministically fails but GLM5.3 solves reveals odd behaviors — kalomaze · 2026-09-17
- Why keep a no-thinking mode in 2026? CoT monitoring and performance both demand reasoning — scaling01 · 2026-09-17
- Dev builds sourced comparison of realtime speech-to-speech models across six providers — mahimairaja · 2026-09-17
- Sergey Karayev tests whether GPT-6 Astra High can one-shot a full GitLab implementation — sergeykarayev · 2026-09-17
- Claude cuts subscription quotas, Codex drops x20 tier as usage curbs pile up — ThePeterMick · 2026-09-17
- Baseten launches safety infrastructure for open models with Hugging Face and Goodfire — baseten · 2026-09-17