OpenAI's incoming Astra model uses a reasoning technique called recurrent depth, also described as opaque recurrence. It processes a query in loops through the model's layers instead of laying out each step as readable text, according to TechCrunch and the-decoder.
The technique boosts performance on math and coding tasks and lets smaller models perform like larger ones, the-decoder reported. The tradeoff is that reasoning shifts from text a human reviewer can read into internal mathematical representations that are effectively invisible, both outlets reported.
That worries the researchers who rely on reading a model's chain of thought to catch misaligned behavior before it causes harm. Redwood Research CEO Buck Shlegeris said that if OpenAI pushes recurrence further, "they'll have the option to massively increase the recurrence and totally destroys CoT monitorability," according to TechCrunch. A May report from the UK's AI Security Institute already warned that opaque reasoning "threatens to severely undermine current oversight methods," the-decoder reported.
OpenAI chief scientist Jakub Pachocki pushed back on the idea that the company is abandoning legible reasoning, telling TechCrunch that "OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models" and calling it a core goal of current research. He also acknowledged the property is "fragile" and "trending in a negative direction," the-decoder reported. Astra is already OpenAI's first model rated at the critical cybersecurity risk tier, having found two previously unknown zero-day vulnerabilities on its own during testing, both outlets reported.
Chain-of-thought monitoring is one of the few tools anyone outside a lab has to check what a reasoning model is actually doing before it acts. If the industry's most capable new model does more of its thinking somewhere nobody can read, builders relying on that model for agentic work are trusting its behavior on the vendor's word, not on evidence they can inspect.