OpenAI’s Astra model uses opaque recurrence technique raising AI safety alarms

OpenAI’s upcoming Astra model will use a reasoning technique called “recurrent depth,” also described as “opaque recurrence,” which allows the model to process queries in loops rather than sequentially, according to reporting by The Information. This approach reduces the legibility of the model’s chain of thought, making it harder to trace the steps the model takes to arrive at an answer.

The technique has drawn criticism from AI safety researchers who argue that it could undermine monitoring capabilities. Redwood Research CEO Buck Shlegeris wrote that if OpenAI expands the use of opaque recurrence, it could “totally destroy” the monitorability of the model’s chain of thought. Longtime safety advocate Zvi Mowshowitz warned that the technique risks enabling a “race to the bottom” among AI labs and suggested that regulation may be necessary to prevent its widespread adoption.

OpenAI has publicly emphasized its commitment to legible chain-of-thought monitoring. Chief scientist Jakub Pachocki stated in a post on X that preserving and utilizing chain-of-thought monitoring has been a core goal of the company’s research program since its earliest reasoning models. He acknowledged that all AI models engage in some opaque reasoning but maintained that chain-of-thought logs remain a valuable tool for monitoring.

Despite these assurances, safety researchers remain concerned. Redwood Research chief scientist Ryan Greenblatt argued that opaque reasoning could scale faster than conventional chain-of-thought methods, potentially pushing models to reason entirely in latent space where their processes are invisible. He expressed hope that OpenAI would halt further development of the technique.

The Information also reported that Anthropic and Google DeepMind are discussing the use of opaque recurrence, suggesting the technique may be gaining traction beyond OpenAI. The emergence of the technique in multiple leading labs has intensified debates about the trade-offs between model capability and the transparency needed for safety oversight.