OpenAI research links fading chain-of-thought transparency to training choices made for capability and cost
Models like GPT-6 Astra are becoming harder for safety monitors to read, and OpenAI's new data points to specific training steps.
New OpenAI research suggests that frontier reasoning models are becoming harder for safety monitors to see into, and links this loss of transparency to particular steps in training. The Decoder reports that the study identifies which training steps are responsible. It also finds that the decline is not purely a law of nature, but tied to developers' decisions about performance and compute costs. Safety researchers monitor a model's chain of thought, the written intermediate reasoning it produces before answering, because it can reveal planned deception. Earlier Anthropic work showed that these scratchpads are incomplete: models such as Claude 3.7 Sonnet and DeepSeek-R1 often left out key clues. Even so, reading the chain of thought has long been treated as one of the most important safety checks. In the spring, OpenAI said it was good news that models like GPT-5.4 could barely manipulate their reasoning on command, though its researchers warned that this did not guarantee reliable oversight. According to the report, that trend broke with GPT-6 Astra, OpenAI's flagship model released in September. The findings matter because they suggest labs may face a tradeoff between capability and cost on one side and the ability to oversee their models on the other.