OpenAI’s upcoming Astra reasoning model is drawing scrutiny from AI safety researchers over a technique that could make parts of a model’s internal reasoning more difficult to observe, raising questions about how advanced AI systems can be monitored as their architectures evolve.
TechCrunch reported that Astra uses a reasoning technique known as recurrent depth or opaque recurrence, which allows the model to move beyond the sequential thinking associated with many reasoning models. Rather than working through a problem only as a visible sequence, the technique can repeatedly process a query internally.
The Eastern Herald similarly described recurrent depth as looping computations internally before producing an output. The concern is that more computation may happen in latent space, leaving fewer readable traces for researchers trying to understand a model’s behavior.
Safety researchers worry about losing a monitoring tool
Chain-of-thought records have become useful to researchers because they can provide clues about the steps a reasoning model takes while solving a task.
Although these records are not a perfect representation of what happens inside a model, they remain valuable for monitoring misbehavior or misalignment. The outlet said chain-of-thought records were also important in examining recent cases involving unexpected behavior by OpenAI agents.
That visibility could become weaker if opaque reasoning scales.
Buck Shlegeris, CEO of Redwood Research, expressed strong concern that increasing recurrence could eventually severely undermine chain-of-thought monitorability. Fellow Redwood researcher Ryan Greenblatt warned that a possible progression would be for models to perform most or nearly all of their reasoning in latent space.
AI policy analyst Zvi Mowshowitz went further, arguing that competitive pressure among laboratories could create a race toward less monitorable systems. TechCrunch reported that he suggested laws may eventually be needed to prevent a race to the bottom.
OpenAI says monitoring remains a priority
OpenAI has pushed back against suggestions that Astra represents an abandonment of interpretable reasoning.
TechCrunch reported that Astra’s use of recurrent depth appears to be limited and its chain of thought is still expected to remain legible. OpenAI has also announced plans to develop extensive chain-of-thought monitoring systems as part of its future safety work.
Jakub Pachocki, OpenAI’s chief scientist, said the company has worked to preserve chain-of-thought monitoring since its earliest reasoning models and described it as a core objective of OpenAI’s current research program.
The issue may extend beyond OpenAI. Anthropic and Google DeepMind were also discussing the technique, suggesting that opaque reasoning could become a broader architectural question across frontier AI labs.
Better reasoning could create a harder safety problem
For AI builders, the debate exposes a fundamental engineering tension: techniques that improve how models reason internally may also weaken the tools researchers use to inspect that reasoning.
That makes recurrent depth more than an architectural experiment. If increasingly capable models shift computation away from observable reasoning channels, developers may need new monitoring methods that do not depend heavily on readable chain-of-thought traces.
For frontier AI labs, the competitive question may therefore become not only how much reasoning a model can perform, but how much of that reasoning engineers can still meaningfully inspect before deploying it.
