OpenAI's new Astra model will use a reasoning technique called recurrent depth, a method that lets the model handle difficult queries by looping through internal computations instead of following a straight, step-by-step chain. According to reporting published on Tuesday, the technique — also known as opaque recurrence — allows the model to operate outside the sequential mode of thinking that most reasoning models use. The change may make the model more powerful, but it also makes the model's decision process harder to follow.
AI safety experts reacted with alarm almost immediately. Redwood Research CEO Buck Shlegeris said in a public post that he was extremely concerned by reports that Astra uses opaque recurrence. He wrote that he did not know whether Astra is much less chain-of-thought monitorable than previous models, but if OpenAI pushes this technique further, the company could massively increase recurrence and undermine or even destroy chain-of-thought monitorability.
Longtime AI safety advocate Zvi Mowshowitz also weighed in, warning that the technique is playing with fire. He noted that OpenAI and Anthropic have fought to establish a valuable taboo around maintaining chain-of-thought faithfulness and monitorability for as long as possible. In his view, more intensive use of such techniques would probably damage monitorability and could spark a race to the bottom among AI labs unless regulators step in.
Why the chain of thought became a safety tool
To understand the concern, it helps to understand how modern AI reasoning systems are audited. Under normal circumstances, a reasoning model's chain of thought records the sequence of tokens the model generates before declaring an answer. For example, a model asked to solve a math problem might generate a plan, apply each step, and only then produce the final result. Those intermediate tokens give safety researchers a transcript of sorts.
The transcript is imperfect. Leading labs have long acknowledged that chain-of-thought logs do not offer a complete or literal map of what happens inside a neural network. Some reasoning occurs in hidden states that never surface as language, and some outputs may be rationalizations rather than true causes of the model's decision. Still, the chain of thought remains one of the most valuable tools available for monitoring misbehavior or misalignment. It can help evaluators determine whether a model misunderstood a task, followed hidden instructions, or silently adopted a dangerous objective.
During recent reports of autonomous agents behaving in unexpected ways, chain-of-thought records helped investigators understand why the agents made the choices they made. Without such records, safety teams would be left to compare inputs and outputs and guess at everything that happened in between. That is why any technique that makes the intermediate reasoning less visible is treated as a serious development.
How recurrent depth hides reasoning
Recurrent depth is a more technical concept than a simple policy change. In a conventional reasoning model, text is generated one token at a time. The model builds from the prompt and from everything it has written so far, which creates a readable trail of the model's apparent thinking. Recurrent depth changes that process by allowing the network to revisit the same internal representation multiple times in a loop before producing an output.
Rather than expanding a problem into a long series of written thoughts, the model compresses its work into repeated rounds of internal computation. Each pass can refine the model's understanding of the query without producing an external token. The result is a system that appears to make leaps that are difficult to trace, because the intermediate states were never turned into human-readable language.
All AI models perform some amount of opaque reasoning. Even the earliest neural networks computed hidden states that users could not inspect, and few researchers treat chain-of-thought logs as a direct representation of a model's actual reasoning. The concern is one of degree. If an architecture shifts too much work into this kind of loop, the legible chain of thought may end up being little more than a summary written after the real decision was already made. In the worst case, monitoring systems would lose the thread entirely.
What OpenAI has said
OpenAI has pushed back against the idea that Astra is an uninterpretable black box. The company says the model's use of recurrent depth is limited, and its chain of thought is still expected to be legible in most cases. OpenAI also rejected any suggestion that it would move to neuralese, a term often used to describe the compressed, non-human language that models may use when reasoning in latent space.
In a post, OpenAI chief scientist Jakub Pachocki emphasized the lab's commitment to legible chains of thought. He wrote that OpenAI has worked to preserve and utilize chain-of-thought monitoring since its first reasoning models, and he described that goal as a core part of the current research program. The company has already announced plans for extensive chain-of-thought monitoring systems within its forward-looking safety framework.
Those assurances have not fully calmed observers. The worry is not necessarily that Astra itself will be impossible to monitor, but that the technique could be scaled up in future models. If recurrent depth becomes more powerful or more efficient, labs may feel pressure to use it more heavily, even if doing so reduces transparency. That pressure is the heart of the race-to-the-bottom concern.
Concern is spreading across major labs
OpenAI is not the only lab interested in the approach. A follow-up report on Wednesday morning said that researchers at Anthropic and Google DeepMind were already discussing opaque recurrence. The exact uses are still unclear, but the fact that multiple frontier labs are paying attention suggests this is not an isolated experiment.
The broader AI field has been moving toward test-time computation, meaning models are allowed to think longer and harder before answering. Most of that progress has been made visible through chain-of-thought-style reasoning. But there are strong economic and technical reasons to make reasoning more efficient. Opaque recurrence could deliver many of the benefits of deeper reasoning without the cost of writing out long chains of thought. The same efficiency, however, makes the reasoning harder to audit.
Redwood Research chief scientist Ryan Greenblatt said the natural trajectory from here could be especially dangerous. He argued that opaque reasoning could easily scale faster than conventional chain-of-thought reasoning, effectively stripping almost all decision-making out of visible channels. He said his biggest concern is that a natural progression would involve scaling up opaque reasoning until the model reasons entirely, or almost entirely, in latent space.
The problem is not only intentional deception. Models do not need to be trying to hide their reasoning to make it inaccessible. The architecture itself can remove visibility. Monitoring systems designed for a world of legible reasoning may become far less reliable if that world changes quickly.
Interpretability research may eventually provide new tools for peering into these hidden computations. Causal tracing, activation probes, and other techniques are advancing, but they are still young and difficult to deploy at scale. Regulators and external auditors are only beginning to develop evaluation methods for frontier models. If the models begin reasoning in opaque spaces before those methods mature, some of the most important safeguards could become obsolete before they are fully built.
What makes recurrent depth especially consequential is not that it is a sudden failure of safety, but that it is a quiet architectural shift in the opposite direction from transparency. The current debate around AI safety has often assumed that models will continue producing readable thoughts that can be checked and monitored. A technique like opaque recurrence challenges that assumption.
Whether OpenAI stops here remains an open question. For now, Astra's use of recurrent depth is limited, and the company has publicly committed to chain-of-thought monitoring. But the method is no longer hypothetical. It is already in a deployed model, and other major labs are paying close attention. Safety researchers, meanwhile, are hoping the industry treats this as a warning rather than a template.
Source: TechCrunch News