The Revolutionary and Dangerous Architecture of OpenAI's Astra: A Journalistic Fabrication

Recently, a loud article circulated online claiming that the new OpenAI model, codenamed Astra, is based on a revolutionary architecture called "recurrent depth" or "looped transformer."

The idea behind the "looped transformer" is that instead of passing a fixed stack of various layers just once, the model passes the hidden state through the same block of layers multiple times before issuing the next token.

Those who follow ML articles already understood at this point that there is no revolution here. This is a well-known architectural trick that traces back to Google's work "Mixture-of-Recursions" from NeurIPS. This idea has been used many times in open source.

Nevertheless, the hype from The Information's article nearly reached the stratosphere. The article also claims that the looped transformer hides the chain-of-thought. Specifically, unlike models that show reasoning in text, the steps in the looped transformer are hidden within the model and thus inaccessible to humans.

Thus, the news quickly transformed from an architectural detail into a debate about AI safety. OpenAI has already been accused of crossing a line. However, the thesis about hidden CoT turns out to be very controversial.

Jakub Pachocki publicly countered, stating that the report is misleading. Sebastian Raschka wrote about it this way:

"Perhaps the journalist from The Information meant some other technique or misunderstood how the looped transformer works.

The mere reuse of layers does not suppress the visible chain-of-thought; it merely adds computations to the hidden states before generating the next token, just like regular transformer layers do.

The only plausible connection that can be drawn is that if the model makes more of these recurrent passes, it may require fewer intermediate reasoning tokens. That is, more computations occur in latent activations that cannot be read as text. But the same effect could be achieved by simply increasing the model size – for example, transitioning from GPT-5.6 Luna to GPT-5.6 Sol.

What conclusion can be drawn, you decide."