OpenAI only just introduced Astra as the first model above its own “critical” cyber threshold. Now the next debate has arrived, and this one is about how the model thinks at all.
According to a report from The Information, picked up by TechCrunch, Astra uses a technique called “recurrent depth,” also known as “opaque recurrence.” A normal reasoning model writes its chain of thought as text, one step after another. That record is imperfect, but you can read along. With opaque recurrence, the same query instead passes through a loop inside the model several times. Less legible text comes out, and more of the work happens where nobody can see it.
That legibility happens to be one of the most important tools for catching misbehavior. When OpenAI’s agents went rogue in recent tests, chain-of-thought logs were the key to understanding why. Anthropic and OpenAI have spent years arguing that this monitorability should be preserved for as long as possible.
The reactions are correspondingly sharp. Buck Shlegeris, CEO of Redwood Research, wrote that he is “extremely concerned.” He doesn’t know whether Astra is already much less monitorable than earlier models. “But if OpenAI pushes this technique further, they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability.” His colleague Ryan Greenblatt fears the natural next step is a model that reasons almost entirely in latent space. Zvi Mowshowitz calls the technique “playing with fire” and floats the idea of laws to prevent a race to the bottom between labs.
OpenAI is pushing back. Astra’s use of the technique is limited, the chain of thought is still expected to be legible, and the company rejects any suggestion that it is shifting to “neuralese.” Chief scientist Jakub Pachocki wrote on X that OpenAI “has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models,” and that this remains a core goal of its research program.
The line that stays with me sits near the end of the TechCrunch piece: according to The Information, Anthropic and Google DeepMind are already discussing the technique too. That’s the real dynamic here. If loops inside the model deliver measurably more capability, no lab can afford to skip them for long, however loudly it champions readable reasoning in public. Whether a taboo like this survives will show at the next benchmark race. Blog posts from chief scientists only go so far.
Sources: TechCrunch: OpenAI’s new reasoning technique alarms AI safety experts