In short
In short: a reasoning model doesn't answer right away. It first writes an internal draft, step by step, called the chain of thought, and only then gives an answer. OpenAI showed the approach with o1 on September 12, 2024: it used reinforcement learning to teach the model to think productively in that draft. The more time it gets to think, the better it does.
Remember solving problems at school. The teacher wanted not just the answer but the working on the page. Not because your handwriting was interesting, but because the working shows where you went wrong and whether you copied. The chain of thought is exactly that working on the page.
The diary that might stop telling the truth
More than 40 researchers, among them Yoshua Bengio and OpenAI's Jakub Pachocki, wrote in a July 2025 paper that the chain of thought offers a rare chance: to read a model's intent to misbehave before it acts. In the same text they warn that this opportunity may turn out to be fragile.
On September 3, 2026 OpenAI released GPT-6 Astra and wrote itself that the model is harder to monitor than its predecessor. It controls its own chain of thought better and is less likely to put incriminating information in it. When pushed on purpose to evade the monitors, it sometimes succeeds. The company found no signs of hidden encoded reasoning.
While the model thinks out loud, you can hear it. The question is whether it will keep thinking out loud.
Why the draft is hidden from the grader
When you read that a lab hides the chain of thought from its automated graders, that isn't red tape. OpenAI says it plainly in its proposals from September 28, 2026: if the grader sees the draft, the model can learn to write drafts that slip past monitoring. You keep the diary honest by not punishing what's written in it.
Primary sources
OpenAI: Learning to reason with LLMs (o1)Chain of Thought Monitorability (arXiv, 2025)OpenAI: GPT-6 Astra safety overviewOpenAI: Towards safety cases for frontier AI trainingThe visual is generated code-art, no third-party images.