Opaque recurrence
What it is
Opaque recurrence — also called recurrent depth — is a reasoning technique in which a model loops the same transformer layers over a hidden internal state multiple times before emitting any output token, instead of reasoning step by step in generated text. Because the extra computation happens inside a latent loop rather than as tokens, it leaves little or no readable trail in the chain-of-thought stream for an outside observer to inspect.
Why it's emerging now
The Information first reported on September 1, 2026 that OpenAI's incoming GPT-6 Astra model made limited use of this technique, and OpenAI's own evaluations reportedly found the resulting reasoning harder to monitor than its predecessor's. TechCrunch and other outlets corroborated the story the next day, and within days The Information reported both Anthropic and Google DeepMind were internally discussing the same technique. Safety researchers — Redwood Research's Buck Shlegeris and Ryan Greenblatt, and commentator Zvi Mowshowitz — quickly published warnings that broader industry adoption could erode chain-of-thought monitorability as a safety tool, turning what had been an academic architecture idea (looped/recurrent-depth transformers) into an urgent, competitively-charged safety debate.
How it relates to what you know
It's a variant of reasoning models / test-time compute that trades away the legibility chain-of-thought prompting normally provides: standard CoT produces an inspectable token trail, while opaque recurrence performs equivalent computation in a hidden loop instead. That makes it a more extreme, safety-relevant cousin of other under-the-hood reasoning changes already on the ledger, like compressed sparse attention and System One models — techniques that also alter how a model "thinks" beneath the visible token stream.