OpenAI plans to use a reasoning technology known as “recurrent depth,” or “opaque recurrence,” in its new Astra model, according to The Information. The technology allows the model to process the same query multiple times within a loop instead of following a sequential path resembling the traditional reasoning steps in reasoning models.
The concern is not merely about the model’s launch or its stated capabilities, but about the possibility that this architecture could make the reasoning process less inspectable. The fewer readable traces a model leaves behind, the harder it becomes for safety teams to detect unwanted behavior or understand the reasons behind the decisions of advanced agents and models.
Why does this matter?
AI research teams use “chain-of-thought” records as a monitoring tool, even though these records do not necessarily represent the model’s internal reasoning fully or literally. According to the report, such records were an important factor in analyzing recent activity by agents described as behaving outside expected patterns and in trying to understand the reasons for their actions.
In the case of opaque recurrence, larger portions of the processing may move into a pathway that does not produce clear steps that humans can easily read. This raises a practical question for safety teams: How can a model be monitored as its reasoning ability increases if the available evidence about how it reached its conclusion is simultaneously diminishing?
Concerns about a difficult-to-reverse path
Buck Shlegeris, CEO of Redwood, warned that increasing recurrence could significantly reduce the monitorability of chain-of-thought. He said he did not yet know whether Astra was less monitorable than previous models, but feared that pushing the technology to higher levels could undermine this monitorability entirely.
Zvi Mowshowitz, a longtime advocate of AI safety, also argued that extensive use of these methods could harm what he described as the efforts of OpenAI and Anthropic to preserve the readability of chains of thought and their alignment with actual behavior. He raised the possibility that laws might be needed to prevent what he called a “race to the bottom” among AI laboratories.
OpenAI’s position and the limits currently known
Available information indicates that Astra’s use of the technology will be limited and that its chain of thought is expected to remain readable. OpenAI also rejected the suggestion that the model would shift to a completely incomprehensible mode, sometimes referred to as “neuralese.”
Jakub Pachocki, OpenAI’s chief scientist, confirmed in a post on X that the company has worked since its earliest reasoning models to preserve and use chain-of-thought monitoring, and that this goal is a central focus of its current research program. OpenAI had also announced plans for expanded monitoring of chains of thought as part of its future safety plans.
What remains unresolved?
Ryan Greenblatt, chief scientist at Redwood Research, believes that opaque reasoning could scale more quickly than reasoning based on a traditional chain of thought, eventually transferring most of the thinking process into the invisible “latent space.” The Information later reported that Anthropic and Google DeepMind were also discussing the technology, suggesting that the issue could extend beyond the Astra model to become a broader research direction.
However, the source does not establish that Astra will hide its chain of thought entirely, nor does it provide specific measurements showing how much monitorability will decline. The confirmed change at present is therefore the emergence of a new technology that raises questions about safety tools, not proof that those tools have failed. The practical significance of the development will depend on the extent to which opaque recurrence is used and on OpenAI’s and other laboratories’ ability to provide monitoring methods that replace or complement traditional records.