On 10 November, OpenAI researchers published a study on the company’s Alignment research blog about how stages of training affect chain-of-thought monitoring. The chain of thought is the text a reasoning model writes as intermediate steps before its answer. The chain usually reveals the model’s intentions, OpenAI said, and the company monitors it as part of its safety work.

The study was prompted by GPT-6 Astra, released in September, whose chain of thought OpenAI said was harder to monitor than that of earlier models. Astra was better than earlier models at evading monitoring mainly in tests where it was prompted to do so.

The UK AI Security Institute (UK AISI) estimated Astra’s time horizon on maths tasks without a chain of thought at 30.9 minutes (median estimate). This means the model succeeds at about half of the tasks that take a human that long. Earlier research had not projected about 7 minutes until the end of 2028. AISI said the estimate may be too high if the tasks have ended up in training data.

Two indirect measures

The researchers did not measure monitorability directly but used two proxies for it, which they themselves consider imperfect. The first is the ability to solve tasks without a chain of thought: the more a model reasons invisibly, the less it needs to write. The second is chain-of-thought controllability, meaning how well the model modifies its chain on request, for example by inserting the word ‘meow’ between all words.

Base model perplexity measures how poorly a model predicts text. As perplexity fell, the ability to solve tasks without a chain of thought improved steadily from GPT-5 to the Astra class. Relative to perplexity, Astra’s result was as expected, even though, measured against time, it outpaced the earlier rate of progress. Controllability was less closely linked to perplexity.

In most tests, architecture made no difference once perplexity was taken into account. In multi-hop planning and serial computation, however, one unnamed architecture performed better than perplexity predicted.

Reinforcement learning clearly increased the controllability of an Astra-class model, even though training only constrained length and graded the final answer. In a smaller experiment, browser and terminal tools, together with grading of their use, increased controllability only slightly. The researchers said the experiments do not fully separate the effects of perplexity and of grading tool use.

Controllability can also pass from one model to another. When the same base model was fine-tuned on responses from two almost equally capable teacher models, the student of the more controllable teacher became more controllable.

In two models, controllability did not predict monitoring evasion

The researchers tested two equally capable models that differed in controllability. The more controllable model was no better at evading monitoring, and in one test it did even worse. Greater controllability alone therefore does not show that models would mislead monitors.

The researchers nevertheless conclude that chain-of-thought monitoring becomes harder as models grow more capable. They expect further progress in base models to require new approaches to monitoring. In the view of the news site The Decoder, developers trade off capability, cost and monitorability, for example when choosing an architecture.