Since GPT-6 Astra launched on September 3, one term has been everywhere that almost nobody outside research circles used before: recurrent depth. A report by The Information put it into circulation, safety researchers sounded the alarm, and ever since, the technique has been treated as the secret behind Astra’s jump in capability. The problem: the technique itself is far less dramatic than its reputation — and the argument worth having is about something else entirely.
Key takeaways
- A looped transformer simply runs the same layers more than once instead of processing them a single time.
- This makes the model more compute-intensive, but not larger — it needs more processing time, not more memory.
- OpenAI has confirmed nothing of the sort for Astra; the attribution comes from a September 2 report by The Information.
- Safety researchers fear the model reasons in ways humans cannot follow. That concern is valid, but it is not specific to this technique.
- An open Chinese model, Nanbeige 4.2, already uses the approach — there you can read the details instead of guessing.
What a looped transformer actually does
A language model consists of stacked layers that process text step by step. Normally, more layers mean a larger model that needs more memory and costs more to run. A looped transformer breaks that equation. It sends the data through the same stack of layers several times instead of adding new ones.
The open model Nanbeige 4.2 shows the principle concretely: it runs a stack of 22 layers twice, behaving roughly like a 44-layer model without having to store the weights a second time. The price is compute, because every pass costs. Its technical report also explains why the team stopped at two passes: more loops added almost no gains while making training considerably slower and more expensive.
AI educator Sebastian Raschka takes a sober view. Reusing layers, he argues, is a small architectural tweak, not a breakthrough. Related ideas have been around for years, among them the NeurIPS method “Mixture-of-Recursions,” where a learned router decides for each token whether it gets one pass, two, or more. Easy parts of a text finish quickly; hard ones receive extra compute.
The case for it: compute where it is needed
The appeal lies in the economics. Many reasoning problems do not require more knowledge but more processing steps — a model has to transform intermediate results repeatedly rather than know more facts. Repeated passes are a cheap lever for exactly that: memory stays flat while reasoning depth grows.
That also explains why research on the topic exploded in 2026. Papers such as “Reasoning with Latent Thoughts” show that comparatively small models can catch up on multi-step tasks when given more passes. For operators this is attractive: a provider can decide per request how much compute it deserves, instead of firing up the same large model every time. How strongly efficiency now drives this market became visible when Google undercut OpenAI sharply on price.
The case against: reasoning nobody can read
The safety objection starts somewhere else. Modern models expose part of their reasoning by emitting intermediate steps as text — the chain of thought. Those steps are currently one of the most important tools for checking whether a model is on the right track or quietly pursuing something else. If a model shifts more work into internal passes, less of it remains readable.
The concern deserves to be taken seriously. It simply does not target this technique in particular. Raschka points out that additional passes by themselves do not suppress a chain of thought — they add computation, just as ordinary layers do. You would get the same effect by simply making the model bigger. Anyone lamenting less legible reasoning is criticizing a general trend, not a specific design. OpenAI’s chief scientist Jakub Pachocki has himself said that preventing unintended harm is getting harder and may become a bottleneck for further progress.
What is established and what is not
A clear separation helps here. It is established that the technique exists, that it works, and that at least one open model uses it. It is not established that Astra uses it. OpenAI has published nothing on the matter; the attribution traces back to The Information and was then picked up by Fortune, TechCrunch, and others. The stronger claim that the method deliberately hides the chain of thought rests on even shakier ground — the original report may be describing a different technique altogether. Anyone working with Astra notices the shift elsewhere anyway: the model responds differently to old prompts than its predecessors did.
What is remarkable, then, is less the architecture than the dynamic around it: a single report about an undocumented detail of a closed model produces a full-blown safety debate within days. That says more about the state of the industry than about loops in neural networks.
Conclusion
Anyone hoping to explain what Astra does better will not find the answer in loops. They are a solid engineering trick that spends compute more efficiently, and their effect is verifiable in open models. The real conflict runs deeper: the more reasoning moves into a model’s internal states, the less can be checked from outside. That shift has been underway for years and is reinforced by every efficiency gain. The honest answer to how much of it happens inside Astra is that nobody outside OpenAI currently knows.