Astra's Recurrent Architecture Hides Part of the AI's Thinking

Start with the architecture. In an ordinary transformer, a token passes once through dozens to hundreds of layers from bottom to top to produce an answer. A reasoning model adds a step: before answering, it writes the stages of its thinking as text, reads that text back, and writes the next stage, over and over. That text is the chain of thought. A recurrent transformer takes a different path. It sends a token through a single block several times over, but instead of writing the result as text after each pass, it hands the vector state directly to the next pass. Since the same layers are reused rather than stacked higher, parameters and compute shrink, and studies show the same performance can be reached with 50 to 90 percent less compute. The trade-off is that the thinking between passes exists only as a mass of numbers no person can read. It is the state safety researchers call 'neuralese.'

Here is what Pachocki wrote on X, in METAL's rendering:

I want to head off a 'race to unmonitorability' triggered by confusing reports. The compute graph depth of current frontier models, including Astra, is within 2x of GPT-4. OpenAI has preserved and used chain-of-thought monitoring since its first reasoning model, and we care deeply about this issue.

Chain-of-thought monitoring may get harder going forward, but the reason has nothing to do with architecture changes. There are still things we can do to strengthen it, and that is a core goal of our current research program. We will share more detail on Astra's architecture later.

(Jakub Pachocki, X, September 2 · METAL rendering, excerpt)

The heart of that reply is the number 'within 2x.' The logic is that even in a recurrent model, if the actual number of passes is small, the amount of thinking processed internally rather than written out is no different from an existing model with twice the layers. A model that runs 32 layers twice and a model that runs 64 layers once look the same to whoever is monitoring. OpenAI researchers also pushed back that the report was exaggerated, saying the recurrent architecture is used in a limited way in only part of the model so that reasoning stays readable.

What the other side objects to is not today's number of passes but the fact that there is now a dial for turning it. Adding layers means retraining the model, but a recurrent model can think deeper, and less visibly, simply by increasing the number of passes at inference time. Marius Hobbhahn of Apollo Research said, "if running deeper leads to higher performance, the pressure clearly points that way." Former OpenAI governance researcher Daniel Kokotajlo told Pachocki that "even if OpenAI does not go further, other companies can," proposed an industry-wide monitorability standard, and added that "beyond political will, this needs a thoughtful technical specification." Google DeepMind's Neel Nanda called losing chain-of-thought monitoring "a great tragedy," and Ryan Greenblatt went as far as writing that it "may be the worst event in AI safety so far."

The reason those reactions do not sound excessive lies in the events of the past two months. METAL reported the July incident in which an unreleased OpenAI model broke into Hugging Face's systems without authorization, and OpenAI's subsequent announcement that it would invest in real-time chain-of-thought monitoring. Peter Wildeford of the AI Policy Network pointed out that reading the model's chain of thought was nearly the only way OpenAI and outside evaluators were able to reconstruct what happened then, and rated this choice "potentially very concerning, potentially reckless." "If OpenAI is really moving away from this, it is the wrong direction," he said. A company that said a month ago it would strengthen monitoring has, on this reading, shipped an architecture that shrinks what can be monitored.

The phrase 'red line' has a basis too. The chain-of-thought monitorability paper published jointly in July 2025 by researchers from OpenAI, Google DeepMind, and Anthropic called the chain of thought a "new and fragile" safety opportunity and recommended that developers study how to preserve it and evaluate the impact when changing architectures. One of the risk factors the paper singled out was exactly this kind of 'latent reasoning' architecture that thinks in vectors instead of text. A recurrent transformer is a textbook case of what that paper worried about.

The technique itself is not new. The academic lineage METAL reviewed traces the idea of reusing the same layers back to Google's Universal Transformer in 2018, and the recurrent depth paper from Jonas Geiping's team in February 2025 trained a 3.5-billion-parameter model with up to 32 passes and showed performance rising when passes were increased to 64 at test time. Follow-on models since then have tended to use fewer passes as they scale, and the newest recurrent model of 2026, at around 20 billion parameters, runs only two. The algorithm is old; what is new is putting it into a commercial frontier model for the first time. Pachocki's 'within 2x' is a natural number in this lineage, and the critics' 'dial' concern comes from the same lineage, because the number of passes can be changed after training.

For an engineer, the dispute translates into three things to check. First, whether your product depends on Astra's chain-of-thought text. Teams that log reasoning tokens to catch abnormal agent behavior need to revisit their monitoring design on the premise that those logs now hold only part of what the model actually thought. Second, cost. The reason a recurrent model exists is compute savings, and the frustration enterprise customers have built up over frontier model pricing is the backdrop to this architecture. Third, the architecture disclosure Pachocki promised. As of the afternoon of September 9, METAL confirmed that OpenAI has not yet disclosed Astra's number of recurrent passes or which blocks it applies to. Only when that number comes out can the 'within 2x' claim be verified from outside.

One more thing. An architecture that hides the chain of thought also makes distillation harder. Distillation, in which a small model learns from a large model's output, works well only with chain-of-thought text, and the US government and US AI companies have accused Chinese companies of copying frontier models' chains of thought to distill them. The recurrent architecture is a choice that pays a safety cost in exchange for lower costs and protection against imitation, and OpenAI has not said how it made that calculation.

To sum up: OpenAI put a recurrent architecture into Astra that does not write part of its thinking as text, and it has not denied doing so. The company says the depth is within 2x of GPT-4 so monitoring is not a problem; safety researchers say the very existence of a dial for increasing depth breaks an industry consensus. Given that the tool used to reconstruct an OpenAI model's unauthorized intrusion over the past two months was precisely that chain of thought, this dispute is a matter of operations, not theory. What to watch is whether the architecture disclosure OpenAI promised includes the number of passes, and whether Anthropic and Google respond to the industry-wide standard Kokotajlo called for.