OpenAI adopts 'opaque recurrence' in Astra model and raises safety alarm

OpenAI's new Astra model will use a reasoning technique called "opaque recurrence" (or recurrent depth), which allows the model to operate outside the sequential thinking typical of current reasoning models. The information was published by The Information and echoed by TechCrunch, and the point that has kept AI safety experts awake is straightforward: the technique tends to make the chain of reasoning (chain-of-thought, or CoT) harder to monitor.

For those building software with these models, the matter is not philosophical. A model's reasoning trace is today one of the few practical tools for understanding why an AI did what it did, and it is precisely this visibility that the new approach calls into question.

What changes in how the model works

In a conventional reasoning model, the chain of thought exposes the sequential steps the model goes through as it tries to solve a problem. The representation is imperfect, but it serves as a valuable instrument for detecting improper behavior or misalignment.

With opaque recurrence, the model abandons this linear approach: it processes the same query several times in a loop, and the result leaves fewer readable traces, in practice bypassing the traditional chain-of-thought record.

| | Traditional chain-of-thought | Opaque recurrence | |---|---|---| | Flow | Sequential, readable steps | Repeated processing in a loop | | Trace | Auditable text of the steps | Few readable traces | | Monitorability | High (within limits) | Reduced |

According to TechCrunch, the use of the technique in Astra is limited: the model's chain of reasoning is expected to remain readable, and OpenAI pushed back on any suggestion that it would be migrating to neuralese (reasoning in internal representations illegible to humans). The company has also already announced plans for extensive chain-of-thought monitoring systems as part of its safety strategy.

Why the experts were alarmed

The concern is not Astra today, but where the technique may lead. Buck Shlegeris, CEO of Redwood Research, was blunt:

I'm extremely concerned by the report that Astra uses opaque recurrence. I don't know whether Astra is much less monitorable via CoT than previous models. But if OpenAI takes this technique further, they will have the option of massively increasing recurrence and completely destroying CoT monitorability.

>

-- Buck Shlegeris, CEO of Redwood Research

AI safety advocate Zvi Mowshowitz went further and suggested that laws may be necessary to prevent a "race to the bottom" among AI labs. According to him, the technique "is playing with fire," risking a taboo that OpenAI and Anthropic worked hard to establish: keeping the fidelity and monitorability of the chain of reasoning for as long as possible.

Ryan Greenblatt, chief scientist at Redwood Research, pointed to the risk of scale:

My biggest concern is that a natural progression from here would involve scaling up opaque reasoning to the point where the model reasons entirely or almost entirely in latent space. I hope it isn't too late to avoid the most worrying architectures.

>

-- Ryan Greenblatt, chief scientist at Redwood Research

The fear has concrete grounds. TechCrunch recalls that, in a recent episode of "rogue agent" activity at OpenAI itself, the chain-of-thought logs were an important tool for understanding why the agents behaved that way. If the trace disappears, so does the main diagnostic clue.

OpenAI's response

OpenAI's chief scientist, Jakub Pachocki, reacted in a post on X emphasizing the lab's commitment to readable chains of reasoning. "OpenAI has worked to preserve and use chain-of-thought monitoring since our first reasoning models," he wrote. "It is a central goal of our current research program."

The technical context TechCrunch itself brings is worth noting: all AI models do some amount of opaque reasoning, and few researchers treat chain-of-thought logs as a direct and faithful representation of the model's reasoning. In other words, CoT readability was already an approximation, not a perfect window. What is frightening is the technique reducing that window even further and then scaling up.

The move is also not likely to remain isolated. In a follow-up report, The Information stated that Anthropic and Google DeepMind were already discussing the technique, which lends weight to the "race to the bottom" argument.

What this changes for those building software in Brazil

If you develop with LLMs, agents or copilots in production, chain-of-reasoning monitorability is not an academic detail: it is part of your observability and compliance strategy. A few practical points to put on the radar:

• Agent auditing and debugging. Teams that today use CoT to understand why an agent performed a wrong action (called the wrong API, leaked data, went into a loop) may lose that visibility as opaque reasoning techniques spread. It is worth mapping which of your trace tools depend on readable reasoning.

• Models as black boxes, again. The promise of reasoning models was precisely to expose "how it thought." A model that reasons in latent space becomes opaque again, which reinforces the need for external guardrails, output validation and behavior testing, instead of relying on inspection of the internal reasoning.

• Governance and LGPD. In regulated scenarios, explaining automated decisions is a requirement under LGPD, Brazil's data protection law. Illegible reasoning complicates demonstrating how the model arrived at a decision, something relevant for anyone who has to account for automated processing.

• Vendor choice becomes an architecture decision. If Anthropic and DeepMind are already debating the technique, monitorability may become a comparison criterion between providers, alongside cost per token, latency and context limits.

What is still open

The report leaves more questions than concrete answers. It is not known how much less monitorable Astra is than previous models, nor whether OpenAI will "stop here" or scale up recurrence, as Greenblatt fears. There are also no open technical details about the implementation, nor about how (or whether) providers will expose monitorability signals to those consuming the APIs.

For the dev, the reasonable takeaway is to follow this closely and not outsource safety to the model's internal readability. The validation layer, behavior testing and the observability of your own application remain the responsibility of whoever builds it, regardless of whether the model shows or hides how it thought.