OpenAI · 2026-09-02 · rumor
Astra's 'recurrent depth' — a reasoning loop that leaves no chain of thought
The Information reports that OpenAI's unreleased Astra model uses 'recurrent depth', looping the same transformer layers instead of writing out each step. Safety researchers say that makes its chain of thought much harder to monitor.
The Information says OpenAI's Astra loops its transformer layers to think, leaving fewer readable reasoning steps behind.
What is it?
Recurrent depth is the technique The Information says OpenAI built into Astra, its unreleased frontier model. Instead of stacking more transformer layers, the model passes its hidden state through the same layers more than once before producing the next token, so part of the thinking happens inside the network rather than as written text. OpenAI has not published the detail itself, so the claim rests on The Information's reporting; an Astra model or system card would confirm or deny it.
How does it work?
A looped transformer reuses one core set of layers repeatedly, which raises effective depth without adding parameters or memory. The trade-off is legibility: the extra passes produce no tokens, so there is no written record of those intermediate steps for a monitor to read. TechCrunch, which covered the story on September 2, 2026, reports that Astra's use of the technique is limited rather than unconstrained.
Why does it matter?
Chain-of-thought text is the main tool safety teams use to catch a model misbehaving mid-task, and it only works while the reasoning stays written down. Redwood Research CEO Buck Shlegeris warned that pushing recurrence further could destroy chain-of-thought monitorability, and AI writer Zvi Mowshowitz said the technique risks a norm that OpenAI and Anthropic had both worked to establish. OpenAI chief scientist Jakub Pachocki replied on X that OpenAI "has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models."
Who is it for?
AI safety and interpretability researchers