“Chain of thought” (Chain of Thought), the term for the written reasoning steps that AI models produce on their way to a conclusion, gives researchers a way to monitor the models’ behavior. However, OpenAI has announced that the written reasoning of its new GPT-6 Astra models is harder to monitor compared to the previous GPT-5.6 Sol model. The reason is that the model produces less written reasoning on some tasks — not that it is deceiving monitoring systems.
The other side of the debate is the “recurrent depth” approach, in which the model performs multiple processing passes over a query but does not write the steps out.
In the wake of these developments, three researchers who want chain-of-thought monitoring work to be preserved:
- Tomek Korbak
- Mikita Balesni
- Jasmine Wang
The three were dismissed from OpenAI on the grounds of violating company policies and sharing sensitive information with an external security organization. OpenAI says the decision is not related to security concerns and argues that monitoring work should continue.
Background
OpenAI is not a new name in the FikirPilot archive: we have published 81 stories mentioning this name in the last 90 days; the most recent is dated October 7, 2026.