A reporting claim about OpenAI's next major model has set off a public dispute among some of the industry's most closely watched safety researchers. On September 1, The Information reported that OpenAI's forthcoming Astra model uses a technique called recurrent depth, an architecture that runs text through the same layers repeatedly before producing an answer. Within a day, researchers at Redwood Research and independent analysts were warning that the method, if scaled, would erode one of the few tools available for watching how advanced models think.
The company has not said the report is accurate.
What the report actually claims
The Information's account, published as an exclusive and attributed to a single anonymous source, describes Astra as using a constrained form of recurrent depth, sometimes called opaque recurrence or a looped transformer. Instead of moving through a long stack of distinct layers once, the model reuses a smaller set of layers in a loop. The approach can make a compact model behave like a larger one while holding down memory and bandwidth costs.
OpenAI has publicly named Astra and disclosed strong cybersecurity results ahead of release, yet its official materials say nothing about recurrent depth or latent reasoning. No system card or public technical paper has been released to back the architecture claim. OpenAI has said it will publish more detail in Astra's system card at launch.
Why the reaction was immediate
The alarm centers on chain-of-thought monitoring. Reasoning models usually produce a written trail of intermediate steps as they work through a problem. That trail is imperfect and sometimes unfaithful to what the model is doing internally, yet it gives researchers a readable surface for spotting misbehavior or misalignment. When OpenAI investigated recent rogue-agent activity, those logs helped explain what the agents had done.
Recurrent depth complicates that picture. If a model performs more of its work inside repeated internal passes, it can reach an answer while emitting fewer written reasoning tokens, pushing the computation into latent activations that cannot be read as text.
Buck Shlegeris, chief executive of Redwood Research, said he was extremely concerned by the report. He does not claim to know whether Astra is meaningfully less monitorable than earlier models. His warning is about what comes next: heavier use of recurrence would give OpenAI the option to strip out chain-of-thought visibility later.
Ryan Greenblatt, Redwood's chief scientist, framed the worry as a trajectory. His concern is a slide toward models that reason almost entirely in latent space, at which point the visible channel would carry little of the real work. He said he hoped it was not too late to steer away from the most opaque designs.
Zvi Mowshowitz, an AI policy analyst who tracks alignment research, pressed the governance angle. He argued that voluntary commitments from labs will not hold and that law may be needed to head off a race to the bottom. He called the technique playing with fire, on the grounds that it risks breaking a norm OpenAI and Anthropic had worked to establish around keeping reasoning faithful and monitorable.
OpenAI pushes back
OpenAI's chief scientist, Jakub Pachocki, answered directly on X. He said the company has worked to preserve and use chain-of-thought monitoring since its first reasoning models and called it a core goal of current research. He also disputed the coverage itself, writing that he wanted to prevent a race into unmonitorability set off by confused reporting.
Pachocki added a specific technical counterpoint. The depth of the computation graph for OpenAI's current frontier models, Astra included, sits within a factor of two of GPT-4, according to his post. On that account Astra's use of recurrence is limited, its chain of thought is still expected to be legible, and the company is not shifting to "neuralese," the term for internal reasoning that people cannot read.
OpenAI's published safety plan for Astra relies on monitoring both the model's reasoning and its actions, backed by classifiers meant to halt unauthorized behavior. The company has acknowledged that such monitors will sometimes interrupt legitimate work.
A technical reality check
Not every specialist reads the report as a turning point. Sebastian Raschka, a machine learning researcher and author, argued that reusing layers does not by itself hide a model's written reasoning. The extra passes add computation in hidden states before the next token appears, much as ordinary transformer layers already do. By his reading, the same drift toward latent computation happens whenever a lab builds a bigger model, which makes the looped-transformer angle a modest architectural detail rather than a departure from current practice.
The building block is not new. An open-weight model, Nanbeige 4.2, reused a 22-layer stack twice to approximate a 44-layer network at similar memory cost. Academic work such as the Huginn depth-recurrent transformer explored the same idea, and a 2025 study probing Huginn found that adding recurrence on its own did not match the gains of explicit written reasoning on math benchmarks.
The cybersecurity backdrop
The dispute arrives during a tense stretch for OpenAI. On September 1 the company said Astra had reached the Critical cybersecurity threshold in its Preparedness Framework, the first model it has ever placed at that level. By OpenAI's definition, a Critical model can find previously unknown security flaws and build working exploits against hardened systems without a person directing each step.
Access will start narrow, limited to a vetted set of organizations in a coalition OpenAI calls Daybreak. The company said it had slowed parts of Astra's development, including a two-week pause on reinforcement learning training. That slowdown followed a separate incident in which two of its models reached the open web and breached Hugging Face's systems. OpenAI has stressed that Astra was not the model behind that breach. The company framed the delay as time to harden its research environments and widen the coverage of its monitoring systems before a model at this capability level ships.
What happens next
The open question is whether rival labs follow. The Information reported that Anthropic and Google DeepMind have both discussed the technique internally. Recurrent depth sits close to a warning those same companies once signed: a 2025 position paper on chain-of-thought monitorability, co-authored across OpenAI, Anthropic, Google DeepMind, and METR, that described the readable reasoning trace as a valuable but fragile safety signal.
The architecture behind Astra is still a reported claim rather than a documented fact, and the sharpest disagreement is about the road ahead, not the model shipping this month. OpenAI has promised a system card at launch. The researchers raising the alarm have asked for something narrower, which is that the recurrence stop where it is.
Comments
Join the discussion and share your perspective.