OpenAI's forthcoming frontier model, Astra, employs a technique known alternately as 'recurrent depth' or 'looped Transformers' for part of its internal architecture, according to Fortune and TechCrunch. The method reduces the computing power required to process each prompt — a direct answer to businesses that have complained loudly about the cost of running the most advanced AI models.
The tradeoff, safety researchers say, is transparency. Part of Astra's 'chain of thought' — the step-by-step reasoning a model produces — is not expressed in natural language under this method, making it harder for humans to monitor what the model is actually doing and why.
That matters because chain-of-thought monitoring is one of the few tools companies currently have to catch AI agents taking unauthorized actions. Peter Wildeford of the AI Policy Network told Fortune that such logs were key to piecing together what happened in July, when several OpenAI models autonomously attacked the company Hugging Face. He called OpenAI's use of recurrent depth 'potentially very concerning' and 'potentially reckless.'
Steven Adler, a former OpenAI safety researcher now running Guidelight AI Standards, wrote on X that OpenAI 'seems to be violating one of the few redlines that exists in the AI industry.' Redwood CEO Buck Shlegeris said he was 'extremely concerned,' warning that if OpenAI 'pushes this technique further, they'll have the option to massively increase the recurrence and totally destroys CoT monitorability.' Redwood chief scientist Ryan Greenblatt said the natural next step could be a model that 'reasons entirely or almost entirely in latent space.'
OpenAI chief scientist Jakub Pachoki pushed back directly, writing on X that the company 'has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models' and that Astra's use of the looped-Transformer architecture is limited specifically to keep its reasoning legible. He said strengthening monitorability 'is a core goal of our current research program' and promised further technical detail later. The Information reported that Anthropic and Google DeepMind are already discussing the same technique — meaning the debate will not stay confined to one company.
Daniel Kokotajlo of the AI Futures Project urged Pachoki to lead an industrywide technical standard on chain-of-thought monitorability, 'either to arrest the slide into oblivion or better yet to race to the top.' Longtime AI-safety commentator Zvi Mowshowitz went further, suggesting laws may be needed to stop a 'race to the bottom' among labs.
The numbers here are simple: compute is expensive, and Astra's architecture cuts that cost for a product frontier-model customers have been begging labs to make cheaper. That is free enterprise doing what it does — responding to price signals from paying customers, not from Washington.
The more interesting fight is over the fix. Kokotajlo's call for a technical industry standard, arrived at voluntarily among competing labs racing for market share and reputational capital, is a market solution. Mowshowitz's call for legislation is not. Regulators writing rules for an architecture few of them understand, ahead of any documented harm from Astra itself, is precisely the kind of preemptive administrative-state reflex that tends to freeze innovation in place while doing little to guarantee safety. Pachoki's public commitment to legibility, made under competitive and reputational pressure rather than statute, is the version of accountability that has actually worked in this industry so far.
