OpenAI’s forthcoming Astra model will reportedly use a reasoning technique called recurrent depth, raising concern that important parts of its internal process could become more difficult to monitor.
The technique, also described as opaque recurrence, allows a model to process the same query repeatedly in a loop instead of relying entirely on the sequential steps associated with conventional reasoning models. The Information first reported Astra’s use of it, according to TechCrunch.
Astra’s implementation is said to be limited, and its chain of thought is still expected to remain legible. Even so, safety researchers warn that a more extensive version of the architecture could shift increasing amounts of reasoning into representations that human reviewers cannot readily inspect.
Recurrence reduces the visible reasoning trail
A conventional reasoning model produces a sequence of intermediate steps as it works toward an answer. Researchers do not treat that record as a perfect or complete representation of what happens inside the model, but it can still help them identify suspicious behavior, errors or signs of misalignment.
TechCrunch notes that chain-of-thought records were useful when investigators examined recent rogue-agent behavior. They provided evidence for reconstructing why agents acted as they did, even if the text did not expose every internal calculation.
Opaque recurrence changes the process by looping over a query several times. That can leave fewer readable traces and may bypass parts of the familiar chain-of-thought record. The concern is not that Astra has already eliminated monitoring, but that the technique creates a path toward doing more reasoning outside visible channels.
Redwood Research CEO Buck Shlegeris said he did not know whether Astra would be substantially less monitorable than earlier models. His warning focused on what could happen if OpenAI sharply increased recurrence in future systems.
Safety advocates fear competitive pressure
AI safety advocate Zvi Mowshowitz argued that the technique risks weakening a norm supported by OpenAI and Anthropic: preserving chain-of-thought faithfulness and monitorability for as long as possible. He suggested legislation might be needed to stop laboratories from competing toward less transparent designs.
Redwood Research chief scientist Ryan Greenblatt identified scaling as his main concern. He said opaque reasoning might improve faster than ordinary chain-of-thought reasoning, potentially moving almost all of a model’s problem-solving into latent space. He urged OpenAI not to proceed toward the most concerning versions of the architecture.
Those statements are warnings about a possible trajectory, not evidence that Astra currently reasons entirely in an inaccessible form. TechCrunch emphasizes that every AI model performs some opaque computation, and few researchers regard visible reasoning logs as literal transcripts of a model’s internal process.
The debate has already spread beyond OpenAI. A follow-up report cited by TechCrunch said Anthropic and Google DeepMind were also discussing the technique. The source does not say that either company has deployed it or committed to doing so.
OpenAI says legibility remains a core goal
OpenAI rejected suggestions that Astra would switch fully to an internal machine language sometimes called “neuralese.” Chief scientist Jakub Pachocki said the company has worked to preserve and use chain-of-thought monitoring since its earliest reasoning models and called that objective central to its current research program.
The company has also announced plans for extensive chain-of-thought monitoring as part of its future safety work. Those commitments provide a stated direction, but the TechCrunch report does not include technical measurements showing how much recurrence Astra uses or how its monitorability compares with previous models.
That missing evidence leaves the disagreement focused on architecture and incentives. OpenAI says the present system retains readable reasoning, while critics worry that a useful performance technique could gradually erode the very signal used to supervise increasingly capable models.
The immediate safety question is therefore not whether chain-of-thought monitoring has disappeared. It is whether recurrent depth begins a transition in which model capability improves by moving reasoning into spaces that monitoring tools cannot follow. Astra’s release and future evaluations will determine how much of that risk is visible in practice.
