Chain-of-thought monitoring has been one of the few reliable tools safety researchers have for understanding what AI models are actually doing. OpenAI’s next model may start eroding that.
According to TechCrunch, OpenAI’s upcoming Astra model uses a reasoning technique called ‘recurrent depth,’ sometimes referred to as ‘opaque recurrence,’ which allows it to process a query multiple times in a loop rather than following the linear, step-by-step thinking that most current reasoning models produce. The result is fewer readable traces of how the model arrived at its output. Less legibility means less oversight, and that’s the part that has safety researchers worried.
In standard reasoning models like OpenAI’s o-series or Anthropic’s Claude, the chain of thought gives researchers and operators a rough window into the model’s problem-solving process. It’s imperfect, but it’s something. When OpenAI dealt with rogue agent behavior recently, chain-of-thought records were central to diagnosing what went wrong. Opaque recurrence, scaled up, would close that window.
Redwood CEO Buck Shlegeris was direct about his concern: ‘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 destroys CoT monitorability.’ Redwood Research chief scientist Ryan Greenblatt went further, warning that opaque reasoning could scale faster than conventional chain-of-thought reasoning, potentially pushing all model reasoning into what researchers call ‘latent space,’ invisible and unreadable by humans.
AI safety commentator Zvi Mowshowitz argued that regulatory pressure may be the only thing that prevents labs from racing each other toward less transparent architectures. The concern is structural: if one lab deploys a faster, more capable model that uses opaque reasoning, competitors face pressure to follow, regardless of the safety tradeoffs.
OpenAI has pushed back, at least partially. Chief scientist Jakub Pachocki posted on X that chain-of-thought monitoring has been a core goal since the company’s first reasoning models, and the company has indicated that Astra’s use of the technique is limited, with its reasoning still expected to be largely legible. OpenAI has also outlined plans for extensive chain-of-thought monitoring systems in its published safety roadmap.
But that may not settle the broader question. The Information reported that both Anthropic and Google DeepMind are already discussing the technique internally. So this is not an OpenAI-only story. It’s a signal about where reasoning model architecture is heading across the industry. And for anyone tracking AI safety seriously, that trajectory matters a lot more than any single model release.
