Hajra Naz
OpenAI’s new Astra model uses a specific reasoning technique. It’s called “recurrent depth.” This approach lets the model operate outside standard sequential thinking. That sequential approach normally characterizes most reasoning models. The Information reported this development on Tuesday. This technique is also known as “opaque recurrence.” According to reports, it will likely make the model’s chain of thought harder to monitor. That possibility has genuinely rattled AI safety experts.
Reports suggest Astra’s use of this technique appears limited. Still, its emergence has raised significant concerns among safety researchers.
“I am extremely concerned by the reporting that Astra uses opaque recurrence,” wrote Buck Shlegeris, CEO of Redwood, in a post after the news broke. “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 destroy CoT monitorability.”
Longtime AI safety advocate Zvi Mowshowitz also weighed in on this issue. He suggested that new laws might become necessary. Specifically, they could help prevent a “race to the bottom” among competing 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.”
Read More: OpenAI Says Astra Is Its First Model to Meet the Critical Cybersecurity Threshold
Under normal conditions, a reasoning model’s chain of thought serves a clear purpose. It provides the sequential steps a model takes while solving a problem. This representation isn’t perfect. Still, it remains a valuable tool. Researchers use it to monitor misbehavior or misalignment. This proved important recently too. During OpenAI’s recent rogue agent incidents, chain-of-thought records helped researchers understand why agents behaved the way they did.
Opaque recurrence works quite differently. The model takes a less linear approach overall. It processes the same query multiple times within a loop. This process leaves fewer legible traces behind. Effectively, it sidesteps the conventional chain-of-thought record entirely.
Importantly, Astra’s use of this technique appears limited in scope. The model’s chain of thought is still expected to remain legible overall. OpenAI has pushed back against certain suggestions too. Specifically, the company denies any shift toward “neuralese.” OpenAI has already announced plans for extensive chain-of-thought monitoring systems. These systems are part of the company’s broader, forward-looking safety plans.
OpenAI chief scientist Jakub Pachocki addressed this concern directly in a post on X. He emphasized the lab’s ongoing commitment to legible chains of thought. “OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models,” Pachocki wrote. “It’s a core goal of our current research program.”
It’s worth noting some important context here. All AI models perform some amount of opaque reasoning already. Few researchers view chain-of-thought logs as a perfectly direct representation of a model’s actual reasoning. Still, these caveats don’t fully dispel growing concern. Opaque recurrence may genuinely make AI reasoning harder to monitor over time.
Read More: Top 6 Types of AI Models Shaping the Future of Technology
This concern grows especially relevant as the technique spreads across different models. In a follow-up report Wednesday morning, The Information shared additional context. Both Anthropic and Google DeepMind were reportedly already discussing this same technique internally.
Ryan Greenblatt, chief scientist at Redwood Research, responded to this news directly. He warned that opaque reasoning could scale much faster than conventional chain-of-thought reasoning. In theory, this could effectively remove reasoning entirely from visible channels.
“My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in latent space,” Greenblatt wrote. “I hope it isn’t too late to avoid the most concerning architectures and that OpenAI will stop here.”
