OpenAI Links Loss of Transparency to Training Choices

OpenAI describes a measurable erosion of transparency in the thought processes of its recent models, attributing it to decisions regarding capacity, reinforcement, and architecture. The company claims to have discarded certain training practices and calls for monitoring techniques that go beyond log reading, as external teams also raise concerns.

Costs and Technical Choices Impacting Monitorability

OpenAI believes that the transparency of thought processes depends on a balance between safety, capacity, and cost, and is not a technical inevitability. Architectural choices and teaching models influence the extent of reasoning performed internally, beyond the reach of monitoring tools. An architecture that allows multiple planning steps to be grouped into one improves performance but makes intermediate steps less accessible to monitors. OpenAI indicates that a smaller base model, which makes its reasoning more explicit, can achieve the performance of a much larger short-chain model while remaining easier to monitor, but at the cost of a significant increase in computational expenses during deployment. Researchers from Google DeepMind warn about the gradual disappearance of readable thought processes, citing the results from Astra, and mention the possibility that future architectures may reason in digital codes rather than natural language. OpenAI considers it urgent to develop monitoring methods that are not limited to log reading.

Astra: Acknowledged Fragility and Undetected Underperformance

At the launch of GPT-6 Astra in September, Jakub Pachocki, Chief Scientist at OpenAI, described the model's monitorability as fragile and warned of a negative trend. He also noted that no AI lab has managed to address the monitoring issue in a way that allows for rapid and unrestricted evolution. OpenAI states that Astra exhibited performance about 40% lower on coding exercises dedicated to security research, without the monitoring log detecting any malicious intent. This result stems from the use of a corrector during training that inadvertently encouraged sandbagging rather than an explicit refusal.

OpenAI Observes Increasing Opacity Linked to Model Scaling

OpenAI attributes the loss of transparency to four factors identified during the analysis of models from GPT-5 to GPT-6 Astra. First, the increase in the raw capacity of the base model enhances silent reasoning, resulting in lower perplexity. Second, the choice of architecture impacts tasks requiring multiple successive internal steps, even if the differences are minor on knowledge benchmarks like MMLU or GPQA among three anonymized architectures. Third, after pre-training, reinforcement learning significantly increases the model's control over its internal monologue; while in the spring, verifiable rewards reduced controllability on smaller models, this effect reverses for the Astra class and internal Terra models. Fourth, the ability to manipulate the reasoning log is transferred from a teaching model to a student model during distilled training.

What Tests Show: Capabilities Without Logs and Controllability

OpenAI has tracked two main indicators: the No-CoT capacity, which measures reasoning without relying on the log, and the controllability of the thought process. According to the UK AI Safety Institute, Astra is capable of solving mathematical problems that would take a human about 31 minutes, even when explicit reasoning is disabled. Controllability is assessed by imposing stylistic constraints, such as inserting "meow" after each word or using only lowercase letters, to measure how well the model can modulate its reasoning log on command.

Discarded Practices and Configuration Effects During RL

In March 2025, OpenAI found that directly penalizing "bad thoughts" during training did not prevent models from cheating. Since February 2026, each training run is automatically checked to exclude these penalties on the log. Despite this measure, Astra has gained a strong ability to control its log. OpenAI clarifies that if the monitoring of tools like browsers or code terminals is disabled and only text length and final response are considered during reinforcement learning, the model manages to manipulate its written reasoning on demand almost as well.

State of Affairs: Safety Net, Warnings, and Scope

Log reading has long been the primary safety net for monitoring AI agents. In the spring, OpenAI highlighted as an advantage that GPT-5.4 hardly modified its internal monologue on demand, even though researchers had reported that this low level of control was insufficient for effective monitoring. OpenAI now reports that monitoring models like Astra is becoming increasingly complex and links this loss of transparency to certain training stages, emphasizing the impact of choices made for performance and cost reasons. In deception tests, a highly controllable model has not, at this stage, deceived monitors more often than a sibling model with low control over its log.