OpenAI CoT-Control: Reasoning Models Can't Hide Their Thinking

OpenAI Blog·high signal

OpenAI published CoT-Control, an open-source evaluation suite measuring how well reasoning models can manipulate their own thought processes. Tested 13 models — controllability scores range from 0.1% to 15.4% for frontier models. GPT-5.4 Thinking achieves just 0.3% controllability at 10K character chains. This low score is framed as good news: models that can't control their CoT also can't obfuscate reasoning to evade safety monitors. CoT controllability is now a standard metric on OpenAI system