OpenAI: what affects chain-of-thought monitorability across the model lifecycle — Astra shows degradation; no-CoT capability tracks perplexity, RL raises controllability

Why it matters

【Key points】

· Release: OpenAI’s Alignment Research Blog posted “What affects chain-of-thought monitorability across the model lifecycle?” on 2026-11-10; @OpenAI reposted it on 11-11.

• · Observation: With GPT-6 Astra, OpenAI reports worse non-adversarial CoT monitorability, stronger adversarial evasion, and much higher no-CoT capability (reasoning in one forward pass with no chain-of-thought or tool tokens).
• These no-CoT results use an unreleased reasoning-off variant.

• · Findings: No-CoT capability rises smoothly as base-model test loss (log perplexity) falls.
• UK AISI had judged Astra off-trend on a no-CoT math time horizon (footnote: 30.9-minute median vs about 7 minutes projected by end-2028)
• OpenAI says it is not off-trend versus perplexity.
• On Astra-class models, CoT controllability still rises substantially in RL with only length constraints and final-answer grading, and fine-tuning on rollouts can transfer a teacher’s controllability.
• In appendix evasion tests, the more controllable model was not more successful.

【Our read】

Background

CoT monitoring is a pillar of OpenAI’s scalable oversight. This post moves “Astra is harder to monitor” from an observation to a breakdown of the training stack.

Comparison

• This is not a new product or model.
• UK AISI’s 30.9-minute figure is external; OpenAI’s “on-trend versus perplexity” does not cancel it.
• The no-CoT numbers use an unreleased reasoning-off checkpoint, so they are not the public Astra’s capability.

Impact

• If latent reasoning and CoT controllability rise with scale and can pass through model-generated data, reading the chain of thought becomes a weaker guarantee.
• A rise in the proxy is not proven intent-hiding.
• The authors say further pretraining will need new monitoring methods.

【Verifiable】

• · OpenAI Alignment blog (dated 2026-11-10)
• · Official @OpenAI repost (2026-11-11)