Astra, the model OpenAI is preparing to reason without leaving a readable trace, worries AI safety researchers

By TechCrunch · September 2, 2026.

OpenAI is preparing its next model, Astra, with a new reasoning technique: "opaque recurrence," as The Information revealed on Tuesday. Today's reasoning models unfold their process in a chain-of-thought (CoT): a sequence of steps in text, in principle readable, that the model generates before giving its final answer. Astra breaks that linear pattern: it processes the same query several times in an internal loop, and the result leaves far fewer traces for a human observer to follow.

The reaction from the AI safety community was immediate. Buck Shlegeris, chief executive of Redwood Research —a research organization focused on the safety and control of advanced models— wrote that he is "extremely concerned" by the news and warned that, if OpenAI decides to take the technique further, it would have the option to "massively increase recurrence and completely destroy" the ability to monitor the chain of thought. Zvi Mowshowitz, a veteran AI safety commentator, went further: he suggested legislation might be needed to halt a "race to the bottom" between labs, and called the technique "playing with fire" against a taboo that OpenAI and Anthropic had cultivated together: keeping their chains of thought faithful and monitorable for as long as possible.

The reason for the alarm is not abstract. The chain of thought, although an imperfect representation of what happens inside the model, is today one of the few practical tools for detecting when a system misbehaves or drifts from what it was asked to do. The article itself recalls that those logs were key to understanding a recent episode of rogue agent activity at OpenAI —the same kind of out-of-control agentic behavior we have already flagged at Zendoric as the most concrete near-term risk of this phase of AI, ahead of any distant superintelligence scenario. Losing legibility in the chain of thought is not a technical nuance: it means giving up a window into why a model did what it did, precisely when those models increasingly operate as autonomous agents with access to real tools.

OpenAI, for its part, plays down the scope of the change. According to the report, the technique's use in Astra is limited, its chain of thought should remain readable, and the company rejected the idea that this amounts to a turn toward "neuralese" —reasoning that would take place entirely in the model's internal representations, never passing through interpretable human language. OpenAI's chief scientist, Jakub Pachocki, responded on X that the company has worked to preserve chain-of-thought monitoring "since its first reasoning models" and that it remains a central goal of its research program. He added, with an important nuance, that all AI models already do some opaque reasoning and that few researchers take those logs as a direct and complete representation of the model's reasoning.

That caveat is true, but it does not dispel the underlying concern: the direction of travel. A follow-up report from The Information, published Wednesday, indicates that both Anthropic and Google DeepMind were already debating the technique. Ryan Greenblatt, chief scientist at Redwood Research, summed up the fear behind all this: that the natural progression from here leads to scaling opaque reasoning until the model reasons entirely or almost entirely in latent space, that is, without ever passing through a form a human can read. "I hope it's not too late to avoid the most concerning architectures," he wrote.

Our read: this episode is a clear example of what the near-term tension between capability and safety that we have been flagging at Zendoric looks like in practice. Opaque recurrence is not bad by design —it can deliver efficiency and reasoning power that the current sequential format does not allow— but every bit of performance gained by sacrificing legibility narrows the room for maneuver of those who have to audit these systems before they operate with greater autonomy over money, infrastructure or sensitive decisions. The fact that two more labs are already looking in the same direction confirms the pattern Mowshowitz points to: without explicit coordination —whether self-imposed or regulatory— one lab's competitive advantage becomes the de facto standard for the entire sector, even if no individual player considers it ideal.

This does not change our underlying thesis about the horizon of this technology: the abundance and medical progress that AI can bring remain an achievable long-term goal. But that future only arrives if, along the way, we retain the ability to understand why models do what they do. The readable chain of thought is, as of today, one of the few governance tools that works without depending on a lab's goodwill. Losing it —even gradually and with good intentions— is the kind of short-term shortcut that can turn an already difficult transition into a far more dangerous one.