If a model can reason in thinking time instead of brain size, the whole scaling game changes. That is the quiet bet buried in the leaks about OpenAI's upcoming Astra: a "recurrent depth" architecture that reuses the same network layer several times before emitting the next token.
It is not a new Transformer that is bigger. It is a Transformer that loops. Instead of standing up 100 separate stations on an assembly line, Astra is said to put one master craftsman on a single workbench and have him re-work the data in place until it is done.
What recurrent depth actually does
Standard LLMs process a prompt in a fixed linear pass: tokens flow down a fixed stack of layers, roughly once, then output the next word. Recurrent depth breaks that. The input loops through a block of fused math components several times, and only then produces a token.
The result is a parameter-reuse mechanism: you trade compute (time) for parameters (space). A small model can behave like a far larger one at inference time, with far lower memory and bandwidth costs.
The technical community has already connected the dots to an open line of research. A 2025 paper (Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach) trained a 3.5B-parameter model and, by increasing internal loop count at test time, effectively shifted its compute budget toward that of a model closer to 50B parameters. The model did not grow; its depth of thinking did.
That trade has a sharp boundary: when extra rounds compound errors instead of paying off, more thinking can make a recurrent model worse. Both sides of the loop matter.
The read: thinking is moving back inside the model
The most interesting part is not the Transformer tweak. It is the direction it shares with latent reasoning work from Meta, Microsoft, and others.
For two years the industry forced models to expose Chain-of-Thought in readable human language. The new wave is the opposite: let the model reason in its own continuous mathematical representation, and skip translating every intermediate step into words. Microsoft's LOTUS, for instance, paired looped transformers with latent reasoning and matched explicit CoT at 3B scale while compressing thinking-phase latency 2.5–6.9×.
The message is structural: we are moving from reasoning that the model narrates to reasoning the model does. Better is not the only reason. Not narrating is also faster and cheaper.
Why this is an industry inflection, not a model launch
Three forces make this consequential beyond one product.
Capital needs a justification. This year Amazon, Microsoft, and Alphabet are committing roughly $600B in datacenter capex, the sharp edge of a broader concentration of AI compute. That spend requires a story in which model quality keeps jumping exponentially. Compute-depth gives exactly that narrative: quality that scales with time and electricity, not just headcount of GPUs.
Competition is real. Anthropic reportedly passed OpenAI in Q2 revenue (≈$11.6B vs ≈$6.7B). OpenAI needs a proof that the "next model" is worth the burn, and its in-house silicon push is part of the same cost equation. Astra's first publicized win is a strong one: its internal version is said to have cracked ten math/CS problems unsolved for over a decade — including the first explicit construction of a non-Sofic group — with proofs formalized as Lean certificates, at a compute cost of roughly $2,000 at current API rates.
Security is the counterweight. Recurrent depth can hide much or all of the reasoning process. The chain of thought that regulators and researchers rely on to audit a model becomes unreadable. This is the crux: the same mechanism that makes models cheaper and smarter makes them harder to inspect, and it lands right as OpenAI's own CEO publicly hit the brakes over Astra being "too capable."
How to read it as a decision
This is not yet official architecture — it is The Information's reporting on insider accounts. Treat the specifics as directional, not final.
- For teams buying inference: watch this as a cost signal. If compute-depth generalizes, the "bigger equals better" procurement heuristic weakens; efficiency vendors and small-foundation-model API bets get more credible.
- For anyone assessing model safety: expect the audit gap to be the next regulatory flashpoint. "Black-box reasoning" will become a first-class consumer concern, and interpretability budgets will need to adapt to opaque internal loops.
- For your mental model of scaling: stop framing progress as "more parameters." The active frontier now cuts across architecture and inference-time compute. The useful question is not how big a model is, but how much it can think before answering.
Recurrent depth is the cleanest illustration yet that the era of the pure parameter curve may be ending — and the era of traded compute-time is beginning.