OpenAI’s new Astra model is reported to use a reasoning technique called recurrent depth, or opaque recurrence, a change that could make its thinking harder to follow. The report landed on Tuesday and immediately drew concern from AI safety researchers who say the method leaves fewer legible traces than ordinary reasoning.
Buck Shlegeris said he was “extremely concerned” by the reporting that Astra uses opaque recurrence. That reaction went straight to the issue at stake: under normal circumstances, a reasoning model’s chain of thought gives a sequential record of how it works through a problem, but opaque recurrence has the model process the same query several times in a loop, which can hide those steps from view.
The concern matters now because OpenAI has already said it plans extensive chain-of-thought monitoring systems as part of its safety work, and its chief scientist, Jakub Pachocki, said the company has worked to preserve and use that monitoring since its first reasoning models. Those records have been treated as a useful tool for spotting misbehavior or misalignment, and in recent cases they helped explain why agents behaved the way they did. That is why even a limited use of opaque recurrence in Astra has drawn attention.
The friction is that OpenAI’s stated commitment and the reported technique point in opposite directions. Zvi Mowshowitz said the method is “playing with fire,” warned that more intensive use would probably damage monitorability, and argued laws may be needed to stop a race to the bottom among AI labs. The question is not whether opaque reasoning exists at all — it is how much Astra uses it, and whether that amount is enough to materially reduce how well the model can be watched.
That gap is where the story now sits. The Information reported on Wednesday morning that Anthropic and Google DeepMind were already discussing the technique, while Ryan Greenblatt warned that opaque reasoning could scale faster than conventional chain-of-thought reasoning and might eventually push a model to reason almost entirely in latent space. For OpenAI, the next test is whether its planned monitoring systems can still read enough of Astra to matter, or whether the move toward opaque recurrence starts to outrun the tools meant to keep it visible.
