OpenAI's Upcoming Astra Model Uses a Reasoning Method That Hides Its Thought Process From Researchers

OpenAI's next flagship model, code-named Astra, will reportedly use a reasoning method that makes its internal thinking harder to read, according to a report from The Information published Tuesday. The technique, called "recurrent depth" or "opaque recurrence," has AI safety researchers openly worried, even as OpenAI insists it's keeping the model's reasoning under control.

Most current reasoning models work in a straight line. They write out a "chain of thought" in plain text, step by step, before landing on an answer. Researchers use those written-out steps to catch a model lying, scheming, or trying to route around safety rules. That record isn't perfect, but it's the main window anyone has into what a model is actually doing.

Recurrent depth works differently. Instead of writing everything out, the model runs the same query through its internal layers multiple times in a loop, refining its answer in a space that isn't natural language. The Information reported that Astra's use of the technique is limited for now, and OpenAI's chain of thought is still expected to be legible. But the door is open for that to change.

Safety Researchers Sound the Alarm

Buck Shlegeris, CEO of Redwood Research, posted his reaction as soon as the report broke. "I am extremely concerned by the reporting that Astra uses opaque recurrence," he wrote. He was careful to note what he doesn't know: "I don't know whether Astra is much less CoT monitorable than previous models." His fear is about where this goes next. "If OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroys CoT monitorability."

Ryan Greenblatt, chief scientist at Redwood Research, went further in comments carried by Digital Today, calling the move toward opaque reasoning potentially "the worst development so far in terms of AI security and safety." He pointed to OpenAI's own investigation of a rogue-agent incident tied to a Hugging Face hack, where chain-of-thought logs were reportedly a key tool for figuring out why the AI agents behaved the way they did. If that visibility shrinks, Greenblatt argued, spotting dangerous behavior gets much harder as models get more capable.

Zvi Mowshowitz, a longtime AI safety commentator, framed the issue as an industry-wide problem, not just an OpenAI one. "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," he wrote, adding that laws might eventually be needed to stop AI labs from racing each other into less transparent, less safe designs.

OpenAI Pushes Back

OpenAI chief scientist Jakub Pachocki responded on X, disputing the more alarming framing of the reporting. He said Astra's computation depth remains close to that of GPT-4 and that preserving legible chain-of-thought "is a core goal of our current research program." Pachocki added that OpenAI has worked to maintain chain-of-thought monitoring since its earliest reasoning models.

According to Digital Today, which cited reporting from The Verge, OpenAI has also delayed parts of Astra's development and launch by several weeks to strengthen safeguards against cyber misuse and unauthorized model behavior. OpenAI said Astra was not directly involved in the Hugging Face hacking incident but that its safety framework reflects lessons learned from it. In a blog post published Tuesday, OpenAI said it will apply additional chain-of-thought monitoring to Astra specifically to catch misaligned behavior early, and that it assessed Astra could carry a dangerous level of cybersecurity capability, including the ability to find and exploit unknown vulnerabilities with minimal human guidance, if given the right tools and access.

OpenAI has not confirmed whether Astra's underlying technical architecture has fundamentally changed from prior models, according to Digital Today's sourcing.

The skeptics deserve a fair hearing here. If a model can genuinely think in loops that never surface as readable text, and labs later crank that recurrence up to squeeze out more performance, the main tool researchers have for catching a model lying or planning something bad simply disappears. That's not a hypothetical Shlegeris and Greenblatt invented. It's the exact mechanism The Information described. Whether Astra crosses that line today is a separate question from whether the industry is building toward it.

On the other side, nobody in this reporting claims Astra is currently unmonitorable. The Information's own account says the technique's use is limited and the model's reasoning is still expected to be legible. Pachocki's response wasn't silence. It was a specific technical rebuttal tied to computation depth, and OpenAI has a stated policy commitment to CoT monitoring going back to its first reasoning models. Multiple outlets covering this story, including KuCoin, Europe Says, Bold News, and Newsbytes App, ran essentially the same account of the same underlying report from The Information without adding independent verification, so the volume of coverage shouldn't be mistaken for multiple confirmations.

The question raised by Mowshowitz is whether individual labs choosing to limit opaque recurrence today is enough, or whether it takes a binding rule to stop the next lab, or the same lab under competitive pressure, from turning the dial up. The Information reported Wednesday that Anthropic and Google DeepMind are also discussing similar recurrent techniques internally, which means this isn't a one-company decision anymore.