The Information reported on Tuesday that OpenAI's new model, Astra, uses a reasoning technique called "recurrent depth," also referred to in reporting as "opaque recurrence." Unlike most reasoning models that follow a largely sequential chain of thought, this method has the model process the same query multiple times in a looped manner.
Why experts are concerned
Critics warn that opaque recurrence could make a model's chain of thought harder to monitor and interpret. Under ordinary circumstances, a reasoning model's chain of thought shows the sequential steps it takes to solve a problem; while imperfect, those records are a useful tool for detecting misbehavior or misalignment. In OpenAI's recent rogue agent incidents, chain-of-thought records helped investigators understand why agents acted as they did.
Buck Shlegeris, CEO of Redwood Research, called the reports that Astra uses opaque recurrence "extremely concerning" in a post after the news broke. Shlegeris said he is unsure whether Astra will be substantially less chain-of-thought monitorable than previous models, but warned that if OpenAI expands the technique, "they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability."
Longtime AI-safety advocate Zvi Mowshowitz likewise said legal measures might be needed to prevent a "race to the bottom" among AI labs. Mowshowitz wrote that the technique "is playing with fire," and that broader use would likely damage monitorability of chains of thought.
Ryan Greenblatt, chief scientist at Redwood Research, responded to the reporting by warning that opaque reasoning could scale faster than conventional chain-of-thought approaches and remove reasoning from visible channels. Greenblatt wrote that a natural progression could be scaling opaque reasoning to the point where a model "reasons entirely or almost entirely in latent space."
OpenAI's response and limits on Astra's use
According to reports, Astra's use of opaque recurrence is limited. OpenAI pushed back against suggestions the model would shift to a so-called "neuralese" — an entirely hidden internal language — and stressed that it still expects legible chains of thought. Jakub Pachocki, OpenAI's chief scientist, wrote on X that "OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models," and called it a core goal of the lab's current research program.
OpenAI has also announced plans to implement extensive chain-of-thought monitoring systems as part of its forward-looking safety measures.
Broader implications
Most AI models perform some degree of opaque reasoning, and many researchers already treat chain-of-thought logs as an imperfect proxy rather than a direct map of internal reasoning. Nevertheless, the concern remains that opaque recurrence could make AI reasoning substantially harder to monitor if its use expands across models and organizations. The Information reported Wednesday morning that Anthropic and Google DeepMind were already discussing the technique.
For now, Astra's application of opaque recurrence appears limited and the model's chains of thought are expected to remain interpretable. Still, safety experts caution that future scaling and architectural choices will be decisive for whether reasoning stays observable or drifts into latent, less accessible spaces.
