The new reasoning model of OpenAI, Astrawill use a technique called recurrent depth. The technique, also known as dull anniversaryallows the model to operate outside the sequential reasoning that characterizes most reasoning models: instead of proceeding in successive steps, the model processes the same request several times in a cycle. The result leaves fewer readable traces and effectively bypasses the conventional register of chain of thought.
The chain of thought of a reasoning model exposes the steps the model takes as it tries to solve a problem. The representation is imperfect, but remains a useful tool for identifying incorrect behavior or misalignments. In the recent story of the OpenAI agents that went out of control, the reasoning chain registers were used to reconstruct why the agents had behaved in that way.
The fear of security researchers
I am extremely concerned by the reporting that Astra uses opaque recurrence. I don’t know whether Astra is much less CoT monitorable than previous models. But if OpenAI pushes this technique further, they’ll have the option to massively increase the recurrence and totally…
— Buck Shlegeris (@bshlgrs) September 2, 2026
The use of the technique in Astra would be anyway limitedbut this was not enough to contain the reactions in the security research community. Buck ShlegerisCEO of Redwood Research, said he was “extremely concerned by reports that Astra uses opaque recurrence,” but distinguished between the upcoming model and the direction it opens. “I don’t know if Astra is significantly less chain-of-thought trackable than previous models. But if OpenAI pushes this technique further, it has the potential to dramatically increase recurrence and destroy chain-of-thought trackability altogether,” he wrote.
Zvi Mowshowitzwho has long been active in the debate on AI safety, added that laws may be needed to prevent one race to the bottom between laboratories. “The technique is a game with fire, and it undermines a taboo that OpenAI and Anthropic have fought to assert, namely that we work hard to maintain fidelity and monitorability of the chain of thought for as long as possible,” he wrote. “More intensive use of such techniques would likely harm monitorability.”
OpenAI’s newest AI, Astra, is reported to use an ‘opaque reasoning’ architecture where more of the reasoning occurs in activations instead of natural language. This may be the single worst development for AI security/safety to date.
The details of Astra aren’t publicly known,… https://t.co/gGalUe06kE— Ryan Greenblatt (@RyanGreenblatt) September 2, 2026
Ryan Greenblattscientific director of Redwood Research, shifted attention to the speed with which opaque reasoning can grow: it could scale more easily than conventional chain of thought reasoning, to the point of removing all reasoning from visible channels. “My biggest concern is that the natural progression from here is to broaden the opaque reasoning to the point where the model reasons entirely or almost entirely in the latent space” he wrote. “I hope it’s not too late to avoid the most worrisome architectures and that OpenAI stops there.”
OpenAI: Astra’s chain of thought remains readable
The model’s chain of thought should remain readable, and the company has rejected the idea of a shift towards “neuralese”. OpenAI has also already announced extensive chain-of-thought monitoring systems as part of its security plans. Jakub Pachockithe company’s chief scientific officer, reiterated the commitment to readable chains of reasoning in a post on
I want to prevent a race into unmonitorability kicked off by confused reporting. The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4.
OpenAI has worked to preserve and utilize chain-of-thought monitoring since…— Jakub Pachocki (@merettm) September 2, 2026
All models perform a share of opaque reasoning, and few researchers read chain of thought logs as a direct image of what the model is doing. Neither of the two clarifications dispels the fear that opaque recurrence makes the reasoning more difficult to monitor, especially if the use extends to different models. According to a later update from The Information, Anthropic And Google DeepMind they are already discussing the technique.
How much the recurrence weighs within Astra, however, is not quantified: the information circulating in the sector speaks of a limited use without indicating the extent, and OpenAI has not currently released technical details on the use of the recurrent depth technique in the new model.
