OpenAI o1 and Hidden Chain-of-Thought
By Satwik ยท May 4, 2026
OpenAI's o1, previewed in September 2024, was the first widely deployed model trained to spend substantial test-time compute "thinking" before answering. Using reinforcement learning to produce long internal chains of reasoning, it posted large gains on math olympiad problems, competitive programming, and graduate-level science questions where prior chat models plateaued.
Why it mattered: o1 reframed the scaling story. Instead of only scaling training, you could scale inference, letting a model deliberate longer for harder problems. This introduced a new axis of capability and cost, and made "reasoning models" a distinct category. It validated the idea that verifiable-reward RL on reasoning traces could teach genuinely better problem-solving, not just better formatting.
The design choice we flag is that o1's raw chain-of-thought was hidden from users; OpenAI showed only a summary. The stated rationale mixed competitive and safety motives, including a desire to monitor unfiltered reasoning for signs of deception or manipulation without the model learning to hide such content to please users. That is a double-edged decision. Hidden reasoning aids safety monitoring in principle, but it also reduces transparency and auditability for the people relying on the answer, and it concentrates trust in the provider.
For our work, o1's reasoning traces are a research asset: monitoring the scratchpad is one of the more promising avenues for catching misaligned or deceptive behavior before it reaches output. But it depends on the traces being faithful to the computation, which is not guaranteed. o1 opened the reasoning-model era and, with it, a live debate about whether we should read, hide, or trust a model's private thoughts.