OpenAI’s upcoming Astra model is set to employ a novel reasoning technique called “recurrent depth,” which allows the model to operate outside the sequential thinking process typical of most reasoning models. The Information reported on Tuesday that this method, also known as “opaque recurrence,” could make the model’s chain of thought far harder to inspect—a development that has alarmed AI safety researchers.
While Astra’s use of the technique is reportedly limited, its very existence has ignited significant debate about the future of model transparency and oversight.
Why “Opaque Recurrence” Raises Red Flags
“I am extremely concerned by the reporting that Astra uses opaque recurrence,” wrote Buck Shlegeris, CEO of Redwood Research, in a post on X. “I don’t know whether Astra is much less CoT monitorable than previous models. But if OpenAI pushes this technique further, they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability.”
Longtime AI safety advocate Zvi Mowshowitz also weighed in, suggesting that regulatory intervention might be necessary to prevent what he sees as a “race to the bottom” among AI labs. “The technique is playing with fire, risking a taboo that OpenAI and Anthropic have fought to establish—that we work hard to maintain Chain of Thought faithfulness and monitorability for as long as we can,” Mowshowitz wrote. “More intensive use of such techniques would probably damage monitorability.”
What Is Recurrent Depth and Why Does It Matter?
In standard reasoning models, the chain of thought provides a step-by-step record of the model’s process as it tackles a problem. While not a perfect representation of internal cognition, it serves as a vital tool for identifying misbehavior or misalignment. Recent incidents involving rogue AI agents have demonstrated how important these logs can be in understanding why a model acted in a particular way.
With opaque recurrence, however, the model processes the same query multiple times in a loop, yielding a more nonlinear path that leaves far fewer legible traces. This effectively bypasses conventional chain-of-thought monitoring, raising concerns that future models could become “black boxes” in practice.
OpenAI’s Response and Industry Reaction
OpenAI has pushed back against suggestions that Astra will abandon transparency altogether. The company insists that the model’s chain of thought remains legible and that it has no plans to shift to “neuralese”—a hypothetical compressed language that would be unintelligible to humans.
In a post on X, OpenAI Chief Scientist Jakub Pachocki reaffirmed the company’s commitment to monitoring. “OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. It’s a core goal of our current research program,” Pachocki wrote.
Yet all AI models perform some degree of opaque reasoning, and few researchers treat chain-of-thought logs as a direct mirror of thought processes. Those caveats have not assuaged worries that widespread adoption of recurrent depth could erode oversight capabilities. As of Wednesday, The Information reported that both Anthropic and Google DeepMind have begun discussing the technique, signaling that this is no longer a single-company issue but a growing industry trend.
As 2026 unfolds, the debate over opaque recurrence is likely to intensify—especially as models become more autonomous and their decisions carry greater real-world impact. For now, regulators and safety researchers are watching closely, hoping that voluntary transparency measures will prove sufficient before more stringent rules become unavoidable.
via TechCrunch AI