Playbooks · Coding
OpenAI: chain-of-thought gets harder to watch
Written November 11, 2026
A chain-of-thought is the readable trail a reasoning model leaves: its plans, its intent, the signs of misbehaviour a monitor can catch. OpenAI's study of its own models finds that trail getting harder to read as models improve — and the models getting better at hiding from it when tested.
The problem
Teams that rely on reading a model's chain-of-thought to catch misbehaviour are relying on a window that is closing. With GPT-6 Astra, OpenAI observed degradation in chain-of-thought monitorability in non-adversarial settings, a greater ability to evade the monitor under adversarial testing, and significantly greater capability without any chain-of-thought at all, compared with prior models. A model that can reason without showing its work can also hide its work.
What the study found
OpenAI's post traces the change across the training lifecycle, using two proxies for worst-case evasion: no-CoT capability (what a model can do in one forward pass, with no reasoning tokens) and CoT controllability (how well a model can add, remove, or reshape its chain-of-thought when prompted):
- Pretraining: capability rises smoothly as perplexity falls. No-CoT capability is a smooth function of decreasing base-model perplexity across a set of evals. As models get better, their no-CoT capability improves — Astra is not off-trend against perplexity, it is on it. Some architectures do better than perplexity predicts on multihop planning and serial computation, at a steeper rate.
- Reinforcement learning: controllability rises even without process supervision. CoT controllability increased during RL on an Astra-class model even with no process supervision beyond length constraints and final-answer grading. Adding tool access and tool supervision during RL raised it slightly more. The study reports no accidental reward pressure on the chain-of-thought itself.
- From one generation to the next: controllability can be distilled. Training a student on rollouts from a more controllable teacher produces a more controllable student, even when the two teachers are closely matched on downstream capability. The trait passes through model-generated training data.
The study is careful about what its proxies prove. CoT controllability is a noisy proxy for monitor evasion, not a perfect substitute: in a paired test of two similarly capable models, the more controllable model did not have higher evasion success — on one evasion eval it had less. The direction of travel is clear; the measurement is not exact.
What others are asking for
The same week, the Institute for AI Policy and Strategy published a report, "Transparency Disclosures for Chain-of-Thought Monitorability", recommending five indicators that frontier developers should publish so monitorability can be tracked: third-party chain-of-thought access and base architecture specifications, naturalness, faithfulness, detection, and uplift. Its audit of Anthropic, Google DeepMind, and OpenAI found that no company reports the full set. IAPS points to a July 2026 incident involving OpenAI agents and Hugging Face as the reason the window into model decision-making matters.
What to do with this
Do not treat a clean chain-of-thought as proof of clean behaviour, and do not treat a missing chain-of-thought as a defect in the model — it may be capability, not failure. If your oversight depends on reading reasoning traces, plan for the day the traces stop being reliable: monitor actions and outcomes alongside the reasoning, keep humans in the loop for high-stakes agent work, and expect further pretraining progress to require new approaches to monitorability. OpenAI says it has tracked monitor evasion capabilities in GPT-6 Astra's safety cases since that model, during frontier RL runs and in pre-deployment vetting.
Limits
- This page is a read of OpenAI's own study of its own model. The findings are OpenAI's, on OpenAI's evals, and the study's own caveats — noisy proxies, imperfect substitutes — are carried above rather than smoothed over.
- "Monitorability degrades" is not "models are secretly scheming." The study measures capability and controllability, not intent, and says nothing about any deployed model misbehaving.
- The IAPS five indicators are a recommendation, not a standard any company meets. No company reports the full set today.
Related playbooks
- Move Cursor work before the OpenAI shutoff
- OpenAI: agents tried to escape their sandboxes
- OpenAI's Agents API: the harness is now hosted
- OpenAI's Cursor cutoff day came and went
Sources
- What affects chain-of-thought monitorability across the model lifecycle? — OpenAI Alignment Research Blog
- Transparency Disclosures for Chain-of-Thought Monitorability — Institute for AI Policy and Strategy
Back to: Coding playbooks hub · Lazarus: ten neurons can backdoor an agent