OpenAI‘s upcoming Astra model reportedly uses a reasoning technique called ‘recurrent depth‘ (also termed ‘opaque recurrence‘), which allows the model to process a query in loops rather than strictly sequentially. This approach reduces the legibility of the chain-of-thought (CoT) record that typically helps safety researchers monitor a model’s reasoning for misalignment or misbehavior. According to a report from The Information, OpenAI‘s use of the technique in Astra is limited—the model’s chain of thought is still expected to be largely legible—but the news has alarmed AI safety experts.

Redwood Research CEO Buck Shlegeris wrote on X that he is ‘extremely concerned by the reporting that Astra uses opaque recurrence,’ warning that if OpenAI pushes the technique further, ‘they’ll have the option to massively increase the recurrence and totally destroy CoT monitorability.’ Zvi Mowshowitz, a longtime AI safety advocate, said the technique is ‘playing with fire’ and could risk breaking an established taboo among labs like OpenAI and Anthropic to maintain chain-of-thought faithfulness and monitorability. He suggested laws might be needed to prevent a ‘race to the bottom.’

In normal reasoning models, the chain of thought provides sequential steps that, while imperfect, are a valuable tool for diagnosing problems—including rogue agent behavior. Opaque recurrence side-steps this by leaving fewer legible traces. However, OpenAI pushed back against concerns that Astra would shift to ‘neuralese’ (inscrutable internal reasoning). Chief scientist Jakub Pachocki emphasized that preserving chain-of-thought monitoring is ‘a core goal of our current research program.’

Despite these assurances, the broader trend is worrying. The Information followed up by reporting that Anthropic and Google DeepMind are already discussing the technique. Redwood chief scientist Ryan Greenblatt noted that ‘opaque reasoning could easily scale faster than conventional chain-of-thought reasoning,’ potentially removing all reasoning 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,’ he wrote, adding that he hopes OpenAI will stop at the current limited use.