OpenAI's upcoming Astra model will employ a reasoning technique called "recurrent depth" that operates outside the sequential thinking typical of most reasoning models, according to reporting by The Information on Tuesday. The technique, also known as "opaque recurrence," is expected to make the model's chain of thought harder to track — a development that has alarmed AI safety researchers. While Astra's implementation of the approach appears constrained, its introduction has nonetheless sparked significant worries among experts who monitor AI alignment.
In opaque recurrence, the model processes the same query multiple times in a loop rather than following a linear path, producing fewer readable traces and effectively bypassing a traditional chain-of-thought record. Typically, a reasoning model's chain of thought offers the sequential steps the model takes while attempting to solve a problem, serving as a valuable mechanism for identifying misbehavior or misalignment. During OpenAI's recent rogue agent incidents, chain-of-thought records proved essential in determining why agents acted as they did. However, Astra's use of the technique appears limited, with the model's chain of thought still anticipated to be readable, and the company has rejected any indication it would transition to "neuralese."
Redwood Research CEO Buck Shlegeris expressed deep alarm in a post following the news, stating "I am extremely concerned by the reporting that Astra uses opaque recurrence." He added that while he's uncertain whether Astra is significantly less monitorable than earlier models, "if OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroys CoT monitorability." AI safety advocate Zvi Mowshowitz wrote that legislation might be required to prevent a "race to the bottom" among AI labs, calling the technique "playing with fire" that risks damaging a taboo OpenAI and Anthropic have worked to establish around maintaining chain-of-thought faithfulness and monitorability. OpenAI chief scientist Jakub Pachocki pushed back, emphasizing the lab's dedication to legible chains of thought and noting the company has already announced plans for extensive chain-of-thought monitoring systems as part of its forward-looking safety plans.
The concern centers on whether opaque recurrence could scale to overshadow conventional monitoring approaches. Redwood Research chief scientist Ryan Greenblatt warned that opaque reasoning might scale faster than traditional chain-of-thought reasoning, potentially shifting all reasoning into invisible channels. His primary worry involves "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." A follow-up report Wednesday morning revealed that both Anthropic and Google DeepMind were already discussing the technique, suggesting the approach could spread across the industry. All AI models perform some amount of opaque reasoning, and few researchers treat chain-of-thought logs as a direct representation of a model's reasoning — but those caveats don't eliminate concerns that opaque recurrence may make AI reasoning more difficult to monitor, particularly as its use expands across different models.
Greenblatt expressed hope "it isn't too late to avoid the most concerning architectures and that OpenAI will stop here," framing the current moment as a decision point for the industry. The question isn't whether Astra itself poses immediate risk, but whether its limited use of opaque recurrence opens the door to more aggressive implementations that could undermine safety monitoring altogether. The tension between advancing model capabilities and preserving oversight mechanisms will likely define how AI labs navigate this trade-off. What remains uncertain is whether competitive pressure or regulatory intervention will ultimately shape whether the industry maintains the monitorability standards safety researchers consider essential.