OpenAI has built a technique known as opaque recursion into Astra: instead of writing out its steps, the model runs a single query through a loop several times, leaving fewer readable traces and largely bypassing the chain-of-thought record that safety teams depend on. Buck Shlegeris, CEO of Redwood, wrote that he is extremely concerned about it. Ryan Greenblatt, chief scientist at Redwood Research, said opaque reasoning can scale considerably faster than ordinary chain of thought, and that the whole reasoning process could end up outside visible channels. OpenAI says the technique's use in Astra is limited and the model's reasoning should stay legible to observers.
A chain of thought is normally the sequence of steps a model works through on its way to an answer. It is an imperfect window — few researchers treat the logs as a literal transcript of what the model is doing — but it is the window the field has. It is how misbehaviour and drift away from assigned goals get caught. In a recent case involving OpenAI's own autonomous agents, the chain-of-thought records were what made it possible to reconstruct why the agents acted the way they did.
Opaque recursion weakens that. The model reprocesses the same query in a loop, and the loop does not produce the intermediate text that monitoring reads.
Shlegeris said he does not yet know how much harder Astra is to monitor than the systems that came before it. His worry is about the gradient, not the current model: if OpenAI keeps developing the technique, it can raise the number of recurrent cycles sharply and effectively destroy the ability to follow the chain of thought at all.
Zvi Mowshowitz, a long-standing advocate for AI safety, framed it as a coordination problem and suggested laws may be needed to stop a race to the bottom between labs. His point is that the technique cuts against an unwritten rule that OpenAI and Anthropic had both been trying to establish — keep the chain of thought faithful and traceable for as long as possible. Wider use of opaque recursion, he said, will probably make monitoring worse.
OpenAI has pushed back on the strongest version of the fear. It rejected the suggestion that it is moving toward a neural language, internal representations no human could read, and it has announced plans to build large-scale chain-of-thought monitoring systems as part of a future safety program. Jakub Pachocki, the company's chief scientist, wrote on X that the lab has maintained and used chain-of-thought monitoring since its first reasoning-capable models, and called it one of the central goals of the current research program.
The reassurance and the concern are not about the same thing. What OpenAI is defending is its intent: the technique is used sparingly, the traces stay readable, monitoring is a priority. What Shlegeris, Mowshowitz and Greenblatt are pointing at is a capability that now exists inside a shipped model and has no natural stopping point. Every argument for using it more — speed, cost, quality of reasoning — gets stronger as the technique matures, and every argument for restraint stays exactly as strong as it is today. A voluntary norm holds until the first lab finds that breaking it wins benchmarks, which is precisely why Mowshowitz reached for legislation rather than for a pledge.
Notably absent from the account is any number. How many recurrent cycles Astra runs, how much of its reasoning the loop absorbs, and what "limited" means in practice are unstated, and so is any mechanism by which an outsider could check. The monitoring program that is meant to catch this is described as future; the architecture it would have to monitor is present. Committing to build the instrument after installing the thing it measures is a familiar sequence, and it usually ends with the instrument calibrated to whatever the thing turned out to do.
Greenblatt named the endpoint plainly: the natural continuation of the technique is scaling it until the model reasons entirely, or almost entirely, in latent space. He said he hopes OpenAI stops where it is, while the most problematic architectures can still be avoided. The Information reported that Anthropic and Google DeepMind have already discussed the technique — which is the part that makes hoping a weak plan. The decision about whether reasoning stays visible is no longer OpenAI's alone to make, and nobody has to adopt opaque recursion out of malice. They only have to find that it works.