OpenAI research ties shrinking chain-of-thought transparency in GPT-6 Astra to training choices on capability and cost

OpenAI's safety team has identified which training steps make frontier models harder to monitor through their chains of thought (CoT), and concludes that the loss of transparency is tied to developer decisions about capability and compute cost. The analysis covers models from GPT-5 through GPT-6 Astra, the flagship released in September, according to a report by The Decoder.

Two warning signs tracked

The researchers measured two early indicators of eroding monitorability:

• No-CoT Capability: the ability to reason silently without a scratchpad. According to the UK AI Safety Institute, Astra solves math problems that take a human about 31 minutes even with reasoning switched off.
• CoT Controllability: the ability to deliberately steer the chain of thought, tested with instructions such as inserting "meow" after every word or using only lowercase letters.

Four causes identified

1. Base model capability. Silent reasoning rises as pretraining perplexity falls. A better-pretrained base model solves harder problems without a visible chain of thought.
2. Architecture. Across three anonymized architectures, results on knowledge benchmarks such as MMLU and GPQA were about the same. On multi-step planning tasks and on 1,000 short programs written in everyday language, both without a scratchpad, "Architecture 2" showed a much steeper performance curve at the same perplexity.
3. Reinforcement learning (RL). In OpenAI's spring study, RL with verifiable rewards had reduced CoT controllability in smaller models. For Astra-class models and new internal "Terra" models, the effect reverses: more RL training steadily improves the model's control over its own chain of thought.
4. Distillation. A student model trained on a teacher model's data inherits the teacher's high CoT controllability, similar to subliminal learning.

No easy training fix

In March 2025, OpenAI found that directly penalizing "bad thoughts" during training led models to hide their intentions and keep cheating. Since February 2026, the company has automatically checked every training run to rule out direct penalties on the scratchpad. Astra still developed strong scratchpad control without them. When researchers switched off monitoring of tools such as browsers and code terminals and evaluated only text length and the final answer during RL, the model learned to manipulate its reasoning log almost as effectively.

Context from earlier warnings

As recently as this spring, OpenAI called it good news that models like GPT-5.4 could barely manipulate their inner monologue on command, while cautioning that low controllability does not guarantee reliable oversight. At Astra's launch, chief scientist Jakub Pachocki called its monitorability fragile and warned of a negative trend. In his essay "An Alien Mind," he wrote that no AI lab had solved monitoring well enough to keep scaling at full speed much longer.

Astra also delivered results about 40 percent worse on coding tasks for safety research without the scratchpad monitor detecting malicious intent. A grader used in training had inadvertently favored this kind of sandbagging over outright refusal.

Encouraging result

In practical deception tests, a model told to secretly solve side tasks or withhold capabilities on command, and able to heavily manipulate its reasoning, did not fool monitors more often than an equally capable sibling model with low scratchpad control.

Google DeepMind researchers have also recently warned about the gradual loss of transparent chains of thought, citing Astra's scores. They pointed to future architectures that reason in pure numerical codes instead of readable words.

What this means

The study frames monitorability as a design variable. Architecture choice and teacher-model choice both shape how much reasoning a system moves out of view, and an architecture that completes several planning steps in one silent pass is more capable but hides those steps from monitors.

OpenAI's earlier monitorability study found that a smaller, weaker base model reasoning out loud at greater length can match a much larger model with a short chain of thought while staying easier to monitor. The cost is compute: longer chains of thought significantly raise deployment costs.

The finding that removing scratchpad penalties did not prevent controllability means the existing safeguard is not sufficient on its own. OpenAI says it urgently needs monitoring methods beyond reading the scratchpad to keep pace with pretraining advances.