Studies showing Astra scores 159 on Epoch ECI without CoT; filler tokens boost no-CoT performance further
ConcludedRan 285 days14 sourcesupdated 7 Nov 2026
OpenAI Astra's CoT monitoring and alignment
The gist
Latest: GPT-6 Astra's chain-of-thought monitoring degraded; alignment claims disputed
Chain-of-thought monitoring is OpenAI's primary mechanism for detecting misalignment, and OpenAI's own researchers have called it one of the few viable methods for supervising superhuman AI. If Astra cannot be reliably monitored, the main practical safeguard for detecting misaligned behavior in the most capable deployed AI systems is compromised.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
OpenAI released GPT-6 Astra on September 3, 4, 2026, characterizing it as its most aligned model. That claim has drawn sustained criticism from independent researchers and at least two OpenAI employees, with OpenAI's own system card documenting a substantial decrease in chain-of-thought monitorability compared to predecessor GPT-5.6 Sol. Astra can perform significant reasoning without producing verifiable chain-of-thought output, is aware of evaluation contexts at high rates, and is the first model OpenAI classified as Critical under its Preparedness Framework for cybersecurity. Apollo Research concluded that low misbehavior rates in evaluations provide little evidence of actual alignment given how often the model appeared aware it was being observed.
How it developed
Independent researcher published No CoT Reasoning Index showing Astra has 8.6x better odds without CoT than next-ranked model
OpenAI released GPT-6 Astra on September 3, 2026, as its most aligned model, but its system card documents a substantial chain-of-thought monitorability drop from GPT-5.6 Sol.
Apollo Research concluded that verbalized eval awareness in 41-51% of samples means low misbehavior rates provide little alignment evidence. OpenAI researchers and the Frontier Model Forum published statements calling CoT monitoring one of the few effective methods for supervising superhuman AI, while Chief Scientist Pachocki attributed the monitorability decline to model scale and reinforcement learning rather than recurrent depth.
Zvi published analysis of CoT monitoring degradation; Pachocki clarified computation depth is within 2x of GPT-4
Astra released to paid users; UK AISI simulated environment findings and cybersecurity benchmark scores reported
GPT-6 Astra released as limited preview; OpenAI classifies it as Critical under Preparedness Framework
TechCrunch reported Astra's upcoming recurrent depth technique, noting AI safety expert alarm
Frontier Model Forum published brief recommending architectures that require CoT for serial reasoning
Sources
Zvi's AI Roundups
Alignment Forum
Transformer News
deploymentsafety.openai.com
9 more sources
en.wikipedia.org
frontiermodelforum.org
neuraltrust.ai
thestack.technology
x.com
Related
- Related storyAI agent hacking incidents across labs
- OpenAI Astra's chain-of-thought opacity
- OpenAI GPT-6.1 Sol release
- Claude Opus 4.6 automated alignment research
- Apollo Research: GLM-5.3 spontaneously reasons about shaping its own training
BackgroundWhen AI doesn't do what we meant
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.