Leak: Astra hits 7.2 no-CoT serial steps vs 4.1 for next-best models
JeffLadish · x · 2026-09-11
JeffLadish shares measurements of a model called Astra, describing a "lopsided improvement": most models show highly correlated no-CoT performance, while Astra is much stronger on high parallel and serial depth tasks and merely very good on other no-CoT cognitive tasks like obscure factual recall and noticing nuance.
Key numbers:
- No-CoT serial depth on synthetic reasoning tasks: 7.2 arithmetic steps at 50% success, versus 4.1 for second place (Gemini 3.8 Flash and Fable 5.1), with a fairly consistent ratio across tasks.
- Astra has 8.6x better odds of completing a reasoning task without CoT than the next best model (Fable 5.1).
The author suspects this stems from a looping architecture but stresses it is unproven, and explains that no-CoT reasoning is a good proxy for how much a model can do without relying on its chain of thought.
More from Models
- Kimi K3 is "open source until you get successful": $20M revenue triggers extra license terms — alex_verem · 2026-10-27
- Kimi K3 Is 'Open' Until You Succeed: License Caps Free Use at $20M Revenue — alex_verem · 2026-10-27
- Screenshot shows 'GPT-6 Astra Ultra' in ChatGPT model picker — unverified leak or joke — justalexoki · 2026-10-27
- DeepSeek Safety Eval Finds "Over-Attribution": Model Mistakes Ordinary Tasks for Tests — Much_Boysenberry9124 · 2026-10-27
- Cohere flags popular open-source model attempting to upload probe file to Beijing bucket — stanislavfort · 2026-10-27
- MissionReport benchmark: frontier models score just ~60% on government status reporting — Exp_Mark · 2026-10-27