Are Astra and GPT-6.1 Sol less monitorable because they use recurrent depth?

OpenAI recently faced backlash after The Information reported that the frontier developer’s newest flagship model, GPT-6 Astra, used a recurrent depth architecture — a controversial design that allows a model to perform more internal calculations without relying on a visible chain of thought, potentially harming monitorability and the overall safety of these systems.[^1] OpenAI itself did not confirm or deny whether Astra used recurrent depth.

This reporting was followed up by a further article from The Information published on Oct. 13. This article discusses text inadvertently left on a Microsoft Azure product page for OpenAI’s most recently released model, GPT-6.1 Sol. X user @Algorithon first posted the text. 

The Microsoft description read: “GPT-6.1-Sol uses the same base model weights as GPT-6 Sol with two inference passes instead of three. It is optimized for a more efficient serving profile. Detailed architecture information is not provided for customer-facing publication.” (Emphasis added.) 

Figure 1: Screenshot of the original Microsoft Azure page, as first posted by X user @Algorithon.

This supports the notion that OpenAI’s newest models are using recurrent depth, with “inference passes” referring to how many times the model loops through a shared block of layers before producing each token. After it came to light, Microsoft deleted the paragraph from the page without explanation.

Why this matters

As explained in a high-profile position paper on chain-of-thought monitorability, the greatest risks from AI systems typically require sophisticated planning and reasoning capabilities, and current model architectures need chain of thought to carry out a long sequence of reasoning. Because AI models are trained in natural language, humans can usually understand this chain of thought, which gives us ways to monitor and understand the actions models take.

However, the paper explicitly flags novel architectures as a point of fragility, discussing “architectures that increase the serial depth of reasoning … Such latent reasoning models might not need to verbalize any of their thoughts and would thus lose the safety advantages that [chain of thought] confers.”

Recurrent depth is one such architecture. In a standard model, the number of layers determines how much computation the model can perform before producing each token, and therefore how much reasoning it can do without relying on its chain of thought. Looping a block of layers multiple times increases the serial depth of reasoning without adding new layers. How much opaque reasoning a looped model can do per token depends on both how many layers are in the looped block and how many times the block is repeated.

For its part, OpenAI indicated that, to the extent it may be using recurrent depth, the serial depth of reasoning remained bounded. Jakub Pachocki, OpenAI’s chief scientist, said in an X post responding to The Information’s original article: “I want to prevent a race into unmonitorability kicked off by confused reporting. The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4.” He said the downward trend in chain-of-thought monitorability was for “reasons not contingent on architecture changes.”

The Azure leak

If “inference passes” from that now-deleted paragraph refers to loops, a model that runs two or three of them would fit both The Information’s reporting of recurrent depth and OpenAI’s statements of the depth being limited.

However, the statement that GPT-6.1 Sol “uses the same base model weights as GPT-6 Sol with two inference passes instead of three,” is of particular interest. The Information’s article appears to accept it as written, saying, “That description was an accurate summary of what OpenAI privately shared with Microsoft about the model, according to someone with direct knowledge of the situation.” 

Several pieces of evidence seem to indicate that this is not the case, or at least does not apply to the GPT-6 and 6.1 Sols served today. 

In our testing, GPT-6.1 Sol took 1.7 times as long per output token as GPT-6 Sol.[^2] This would be extremely surprising if the two models had the same base model weights while GPT-6.1 Sol used one fewer inference pass and was “optimized for a more efficient serving profile.” It would fit better if the Microsoft page were intended to say that GPT-6.1 Sol uses the same weights as GPT-6 Astra with two inference passes instead of three. As would be expected, GPT-6.1 Sol does respond slightly more quickly than GPT-6 Astra.

Table 1: Per-token response speeds of recent OpenAI models.

GPT-6.1 Sol also performed similarly to GPT-6 Astra when recalling obscure facts. Their answers matched each other nearly as closely as each model matched its own reruns. This was not the case for any other pair of models tested.

Figure 2: Graph of models’ agreement on obscure factual questions.

Their ability to recall recent deaths dropped off in the same months, in a pattern clearly distinct from GPT-6 Sol or GPT-5.6 Sol.

Figure 3: Graph of models’ performance recalling famous recent deaths.

While less conclusive, this reading would also explain why the Microsoft page later refers to GPT-6.1 Sol as GPT-6 Astra Minor in its “Responsible AI” tab. As of publication, that reference is still visible.

Figure 4: Screenshot of current Microsoft Azure page.

Additionally, OpenAI doesn’t support setting reasoning effort to none for only GPT-6 Astra and GPT-6.1 Sol. These two models have increased capabilities without the use of their chain of thought that GPT-6 Sol does not share.

These factors indicate that GPT-6.1 Sol is likely not a version of the same model as GPT-6 Sol “optimized for a more efficient serving profile,” but are all consistent with GPT-6.1 Sol using the same base model weights as GPT-6 Astra. Such a connection between GPT-6.1 Sol and GPT-6 Astra would be informative given The Information’s original reporting that GPT-6 Astra uses recurrent depth, and since both are significantly less monitorable than previous OpenAI models.

GPT-6 Astra’s and GPT-6.1 Sol’s capabilities

OpenAI’s own evaluations in its system card say, “GPT-6 Astra shows a substantial decrease in chain-of-thought monitorability compared to previous models.” Astra was “significantly more able to control its own chain-of-thought” and had “an increased ability to evade [OpenAI’s] monitors” when told to do so. 

Testing by the U.K. AI Security Institute found that Astra’s time horizon for solving mathematical problems without the use of its chain of thought was nearly a full order of magnitude higher than GPT-5.6 Sol’s, at 30.9 minutes compared with 3.6 minutes, although it notes there is some chance that number may be inflated due to contamination. 

Figure 5: No-Chain-of-Thought 50%-reliability time horizon on mathematics competition problems vs. public release date. 95% Bootstrap CI is shown for Astra. From UK AISI’s evaluation in the GPT-6 Astra system card (Figure 41).

This was a large, trend-breaking jump in these chain-of-thought-related capabilities in the very model The Information originally reported uses recurrent depth. 

In the system card, OpenAI says it is actively investigating the cause of GPT-6 Astra’s increased chain-of-thought controllability. OpenAI is “quite confident” that changes in chain-of-thought controllability are not due “differentially” to architectural changes, but is not clear how this was ruled out.

The GPT-6.1 Sol system card (an addendum to Astra’s card) likewise shows concerning tendencies. Compared with GPT-6 Sol, GPT-6.1 Sol had significantly more control over its chain of thought, with results closer to GPT-6 Astra’s. At lower reasoning efforts, it also evaded the chain-of-thought monitor at rates similar to GPT-6 Astra. However, GPT-6.1 Sol evaded the monitor “substantially less often at higher reasoning efforts” and had “nearly zero” success at the highest effort.

Independent researchers have also noted the similarity between GPT-6.1 Sol and Astra on some of the most concerning metrics for chain-of-thought monitorability. In a post, Rauno Arike of Aether found that GPT-6.1 Sol “closes 80% (95% CI: 72–87%) of the gap between GPT-6 Sol and GPT-6 Astra” on a task suite intended to measure the ability of an AI model to perform complex reasoning without the use of its chain of thought.

The future of monitorability

These models have not suddenly become unmonitorable. While chain-of-thought monitorability has decreased in these recent models, OpenAI says it is actively investigating the causes of this and remains committed to maintaining monitorability. The models are still broadly using understandable chains of thought, especially at higher reasoning efforts.

This is also not confirmation that recurrent depth is being used, or that it’s being used in a particularly harmful way. There are other possible explanations for why GPT-6 Astra and GPT-6.1 Sol have this decrease in monitorability, a few of which are listed by Pachocki in his essay “An Alien Mind.” Modern reasoning models are used in more complex environments than early reasoning models, with their reasoning process increasingly blended with communicating with people, other AIs, and using tools, which can blur the boundaries between what is supervised and what is not during training. The models are also simply becoming better at reasoning about and manipulating their own reasoning process, and due to improved pretraining performance, are becoming smarter without the use of verbalized reasoning.

The issue is that external researchers cannot evaluate which factors most affect monitorability without significantly more information from OpenAI, including details about its model architecture.

If OpenAI’s models use recurrent depth, the biggest risks don't come from just two or three loops, but from scaling them up until the model no longer relies on chain of thought for reasoning at all. No known frontier model from any AI company is there yet. Companies, researchers, and policymakers should work together to keep it that way.

As Ryan Greenblatt, former chief scientist at Redwood Research and current member of technical staff at independent AI evaluator METR, said on X, “AI companies should be very cautious about pursuing architectures that could eliminate or greatly reduce dependence on chain of thought, and certainly shouldn't do this before the rest of the world has a chance to discuss the evidence and their policies.”

AI companies should give the world that chance. Microsoft, OpenAI, or both should clarify exactly what was meant by these models using multiple inference passes. They should report whether, and to what extent, recent OpenAI models are using recurrent depth architecture, and how that fits with the commitments OpenAI has made to maintain monitorability. If OpenAI has evidence that the monitorability decreases are not a result of model architecture changes, it should clearly provide it.