OpenAI's Astra model reportedly uses recurrent depth, raising concerns about chain-of-thought monitorability

The core change is architectural: Astra uses recurrent depth, also called opaque recurrence, which processes a query multiple times in a loop rather than producing a linear chain of thought. This leaves fewer legible traces, meaning the conventional chain-of-thought record that safety researchers use to inspect model behavior becomes harder to read. The technique is reportedly limited in Astra, and OpenAI says the chain of thought should still be legible, but the concern is about where the technique could go if scaled.

The cost of adoption is not measured in dollars here but in observability. Chain-of-thought logs were an important tool in diagnosing OpenAI's recent rogue agent activity, and if opaque recurrence scales up, that diagnostic channel could degrade or disappear. Ryan Greenblatt's concern is that a natural progression would push reasoning entirely into latent space, removing it from visible channels altogether. Buck Shlegeris framed it as giving OpenAI the option to massively increase recurrence and destroy chain-of-thought monitorability.

OpenAI has pushed back against the strongest version of the concern. Jakub Pachocki emphasized the lab's commitment to legible chains of thought and pointed to announced plans for extensive monitoring systems. The company reportedly pushed back against any suggestion it would shift to neuralese. This is a hedge, not a guarantee: the technique exists in Astra, and the company's stated commitment does not foreclose future expansion.

The story is framed almost entirely through the lens of safety expert alarm, which is the angle TechCrunch chose. The material notes that all AI models already do some opaque reasoning and that few researchers treat chain-of-thought logs as a direct representation of reasoning. That context matters because it means the boundary being debated is one of degree, not kind. Zvi Mowshowitz argued the technique risks a taboo that OpenAI and Anthropic previously fought to establish around chain-of-thought faithfulness.

Where it stops working is clear from the material: if opaque recurrence scales to dominate a model's reasoning, chain-of-thought monitoring becomes ineffective. The Information reportedly noted that Anthropic and Google DeepMind were already discussing the technique, which suggests the concern is not isolated to one lab. With only one feed carrying this story, the broader industry response is thinly sourced, and the reported discussions at other labs are attributed to a follow-up report rather than direct statements.