When Models Stop Thinking Out Loud

October 4, 2026

One thing that stood out to me from the release of GPT-6 Astra is the reported¹ use of opaque recurrence. This was a new concept to me, so I’ll explain what I learned using an analogy: imagine a (simplified) chef preparing a meal.

We can break down the capabilities of our simple chef into three dimensions²:

1. Width: Their experience and knowledge of cooking (the model parameters).
2. Length: The number of steps they can write down and work on while they cook (reasoning tokens).
3. Depth: The ability to taste test while they are cooking, think about what’s missing, and refine before the next step (recurrent computation).

The taste and iterate step is the rough idea behind opaque recurrence. The model can repeatedly process and update its internal state before it produces the next token.

The catch is the opaque part of it. With this latent form of reasoning, there is intermediate computation happening internally without producing readable reasoning tokens that we can see. We’re effectively losing visibility into part of the reasoning process.

Why observability matters

One of the more useful tools researchers have right now for monitoring model behavior is chain-of-thought (CoT) monitoring. Rather than just looking at the final output, they can examine the model’s intermediate reasoning to identify potentially unwanted behavior.

It’s essentially the model’s written thought process. If you ask a reasoning model to solve a problem, it might first work through several steps in text before producing an answer. Researchers can use another model to examine those steps and look for signs of unwanted behavior.³

In research published in March 2025, OpenAI showed how monitoring an agent’s chain of thought could help detect reward hacking during reinforcement learning. For example, a coding agent tasked with making tests pass might find out it can modify the test framework itself to skip failing tests, instead of actually fixing the code. Monitoring the agent’s reasoning could help identify when it was trying to exploit the task.

This same research discovered something else I found quite interesting. When the team penalized models for expressing reward hacking intentions, the models could learn to conceal those intentions while still exploiting the tasks.

So even when we have access to a model’s reasoning, that visibility can be fragile.

What happens when reasoning becomes less visible?

Back to Astra. In its system card, OpenAI reported it had decreased chain-of-thought monitorability compared with its predecessor. At the same time, they found the model was better aligned overall.

It’s worth pointing out that OpenAI hasn’t established opaque recurrence as the cause of this regression. Their research suggests it may have more to do with how the model was trained, rather than changes to its architecture. They say they’re “quite confident that changes in CoT controllability are not differentially due to any architectural changes,” and that they’re still investigating the causes. Astra’s use of the technique is also reportedly limited, and OpenAI has said its chain of thought is still expected to be legible.

Interestingly, the same system card found that Astra could solve harder problems without producing any CoT at all. So we’re already seeing models become more capable without necessarily giving us more insight into their reasoning.

And I think that’s the broader concern. If more of a model’s reasoning happens outside of human-readable CoT, we could be losing access to one of the signals we rely on to identify problems during training and deployment.

We’re very early in understanding how models work internally. There’s so much we can’t explain about how they learn and make decisions, and the tools we have to monitor their behavior are already limited.

I don’t think opaque recurrence itself is necessarily the problem. What concerns me more is what comes next. If a new technique makes models significantly more capable, but also harder to monitor, what’s stopping every lab from adopting it to stay competitive? And what happens when the next breakthrough introduces an even bigger tradeoff?

I think we’re going to face increasingly difficult decisions about how much visibility we’re willing to give up in exchange for capability. And I’m not sure we have a good way of making those decisions yet.