OpenAI's Astra reportedly reasons via 'opaque recurrence,' alarming safety researchers

OpenAI's Astra and recurrent depth

According to TechCrunch, drawing on reporting from The Information, OpenAI's upcoming Astra model will use a reasoning approach known as "recurrent depth," also referred to as "opaque recurrence." Instead of moving through a problem in the linear, step-by-step manner that defines most current reasoning models, the model processes the same query multiple times in a loop. The Information's report, published Tuesday, indicated that Astra's use of the technique is limited — but its arrival has nonetheless unsettled AI safety researchers.

What chain of thought is for

A reasoning model's chain of thought lays out the sequential steps it takes while working toward an answer. TechCrunch describes the record as imperfect, but it remains one of the more practical instruments for detecting misalignment or misbehavior: when OpenAI recently investigated rogue behavior by its agents, chain-of-thought logs were a key source of insight into why the agents acted as they did.

Opaque recurrence works around that visibility. By cycling through a query repeatedly rather than writing everything out in order, the model leaves fewer legible traces and effectively sidesteps a conventional chain-of-thought record.

Warnings from the safety community

Buck Shlegeris, CEO of Redwood Research, wrote that he was "extremely concerned" by the reporting and warned about where it could lead. "I don't know whether Astra is much less CoT monitorable than previous models," he wrote, "but if OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroy CoT monitorability."

Zvi Mowshowitz, a longtime AI safety advocate, argued the technique is "playing with fire" because it risks a taboo that OpenAI and Anthropic have fought to establish around keeping chain-of-thought faithful and monitorable. He suggested legislation might be needed to prevent a race to the bottom among AI labs.

Ryan Greenblatt, chief scientist at Redwood Research, focused on the trajectory. His biggest concern, he wrote, is a natural progression in which opaque reasoning scales up until the model reasons "entirely or almost entirely in latent space," pulling reasoning out of visible channels altogether. He added that he hopes it is not too late to avoid the most concerning architectures.

OpenAI's response

For now, Astra's use of the technique appears constrained. TechCrunch reports that the model's chain of thought is still expected to be legible, and OpenAI pushed back on any suggestion that it would shift to "neuralese," reasoning done in internal representations rather than text. The company has also announced plans for extensive chain-of-thought monitoring systems as part of its forward-looking safety work.

Jakub Pachocki, OpenAI's chief scientist, addressed the concerns in a post on X. "OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models," he wrote, adding that it is "a core goal of our current research program."

There are broader caveats as well. As TechCrunch notes, every AI model performs some amount of opaque reasoning, and few researchers treat chain-of-thought logs as a direct readout of a model's internal computation. Still, in a follow-up report on Wednesday, The Information said both Anthropic and Google DeepMind were already discussing the technique — a signal that recurrence could spread across frontier labs.

Why it matters

Chain-of-thought monitoring is one of the few oversight tools that currently works at scale. If recurrent-depth techniques spread and reasoning migrates into latent space, the industry loses much of its ability to audit frontier models at precisely the moment those models become more capable and more autonomous. The debate over Astra is therefore less about a single model's design than about whether chain-of-thought legibility remains a shared norm that labs hold themselves to — or a design choice each lab can trade away for performance.