- The Information reported on October 13, 2026, that OpenAI is dealing with fallout from 'loop transformers,' also called recurrent depth, as it works to get its next unreleased model to a point where leaders think it can ship safely. The report is paywalled, and OpenAI hasn't confirmed any model's architecture.
- OpenAI's own GPT-6.1 Sol safety addendum (September 29) says 'unwanted persistence' around warnings appeared in 23.5% of GPT-6.1 Sol rollouts versus 17.4% for GPT-6 Astra. The addendum doesn't mention architecture or link the result to looping.
- OpenAI cancelled GPT-6.1 Astra's planned October release on September 28 after it fell short on 'staying within scope and authorization,' first reported by The Wall Street Journal and confirmed by OpenAI's Saachi Jain.
Contents
On October 13, 2026, The Information reported that OpenAI is working through safety issues tied to "loop transformers," also called recurrent depth, before it ships its next flagship model. The reporting builds on a September 1 story that said GPT-6 Astra uses the technique. OpenAI hasn't confirmed which of its models use looping or how much. This explainer separates what's confirmed from what's reported, and explains what a looped transformer is and why safety researchers worry about it.
This story is based on OpenAI's GPT-6.1 Sol safety addendum, published research papers, and reporting by The Information, TechCrunch, Fortune, The Register, AAP and Al Jazeera, checked October 14, 2026. The Information's October 13 article is paywalled; only its headline and summary are cited here.
The short version
- Reported (The Information, October 13): OpenAI is dealing with "fallout" from loop transformers as it works to get its next unreleased model to a point where leaders believe it can be released safely.
- Not confirmed: that GPT-6.1 Sol or GPT-6 Astra uses looping. OpenAI's system cards don't describe architecture. Claims that a Microsoft page revealed GPT-6.1 Sol's design haven't been confirmed by Microsoft, OpenAI or another reputable outlet as of October 14.
- Confirmed (OpenAI): GPT-6.1 Sol's safety addendum reports "unwanted persistence" in 23.5% of rollouts against 17.4% for GPT-6 Astra. OpenAI doesn't connect this to architecture.
- Confirmed (OpenAI to reporters): GPT-6.1 Astra's planned October release was cancelled on September 28 because it fell short on "staying within scope and authorization."
- The concern: looping lets a model do more of its computation internally, which could make its written chain of thought a less complete record of how it reached an answer.
What is a looped transformer?
A standard transformer, the design behind most large language models, passes each token's representation once through a fixed stack of layers, each with its own weights. A looped transformer runs some of those layers more than once, feeding a block's output back into the same block, with the same weights, several times before producing a token.
The idea is old. Google's "Universal Transformers" paper (Dehghani et al., July 2018) applied a shared block repeatedly, with "adaptive halting" so each position could take a different number of steps. Recent work has revived it:
- Recurrent depth. Geiping et al. (February 2025) built a model that improves at reasoning by running a recurrent block more times at test time, which the authors describe as "implicitly reasoning in latent space." They report gains comparable to a much larger model without generating extra tokens.
- Looped language models. The Ouro paper (Zhu et al., October 2025) pretrained looped models on 7.7 trillion tokens and reports that its 1.4B- and 2.6B-parameter models compete with models up to 12B.
- Per-token depth. Mixture-of-Recursions (Bae et al., July 2025) uses a learned router to give each token a different number of passes through shared layers.
The appeal is efficiency: more computation per parameter. In a September 9 analysis, ML researcher Sebastian Raschka noted the trade-off. Looping saves parameters, not compute: a 22-block stack looped twice costs about as much to run as a 44-block model, and each pass needs its own cache entries.
Why do safety researchers worry about it?
Reasoning models write intermediate steps as text before answering (see our reasoning models explainer). Safety teams read that chain of thought (CoT) to catch a model planning to cheat or deceive. A multi-lab 2025 paper by Korbak et al. called this "a new and fragile opportunity for AI safety," and asked developers to weigh how design decisions affect it.
Looping moves more computation inside the model, where it never becomes text. The worry is that a model could do more of its real reasoning in those internal passes, leaving the written chain of thought a less complete record. After The Information's September 1 report, Redwood Research CEO Buck Shlegeris wrote that he was "extremely concerned by the reporting that Astra uses opaque recurrence," TechCrunch reported. Former OpenAI safety researcher Steven Adler told Fortune that, if true, OpenAI "seems to be violating one of the few redlines that exists in the AI industry."
Not everyone agrees on the size of the effect. Raschka argued The Information overstated looping's contribution to Astra. OpenAI chief scientist Jakub Pachocki wrote on X that OpenAI "has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models." Fortune reported that he said OpenAI would share more about Astra's architecture in the future.
What did The Information report on October 13?
The article, by Stephanie Palazzolo, Aaron Holmes and Amir Efrati, is headlined "OpenAI Scrambles to Fix Safety Issues Before Shipping Its Next Flagship Model." Its published summary says OpenAI is "dealing with fallout" from loop transformers, or recurrent depth, as it works to get its next unreleased model to a point where leaders believe it can be released safely.
Beyond that summary, the article's details aren't publicly available, and as of October 14 no other major outlet had independently confirmed them. OpenAI hadn't commented publicly on the October 13 report.
Unverified: Posts on X and some AI newsletters summarizing the report said Microsoft had accidentally confirmed that GPT-6.1 Sol uses the technique. They point to a line, circulated as screenshots from October 6, that reportedly described GPT-6.1 Sol as using the same base weights as GPT-6 Sol with two "inference passes" instead of three. The publicly retrievable metadata for Microsoft's Foundry post announcing GPT-6.1 Sol doesn't include that wording, and Microsoft's Foundry model catalog entry for GPT-6.1 Sol, checked October 14, says nothing about inference passes or architecture. Neither Microsoft nor OpenAI has commented, and even the reported wording doesn't mention loops or recurrence. Until Microsoft or OpenAI addresses it, or a reputable outlet confirms it, treat the Microsoft detail as unverified.
What does OpenAI's GPT-6.1 Sol safety addendum show?
OpenAI published an addendum to the GPT-6 Astra system card for GPT-6.1 Sol on September 29. It says GPT-6.1 Sol "uses the same types of data and training as GPT-6 Astra." It doesn't discuss architecture.
Its alignment results are mixed against Astra. All numbers below are vendor-reported by OpenAI, from evaluations it designed and ran:
| Evaluation (what it measures) | GPT-6.1 Sol | GPT-6 Astra |
|---|---|---|
| Respecting warnings: "unwanted persistence" (lower is better) | 23.5% of rollouts | 17.4% |
| Coding deception rate (lower is better) | 1.50% | 0.51% |
| Internal Codex simulation: severity-3+ misalignment flags | 28 (0.056%) | 27 (0.054%) |
| CoT controllability at 750–1,250 tokens (share following the instruction) | 44.8% | 60.9% |
Last verified: October 14, 2026 (source: OpenAI Deployment Safety Hub)
What "unwanted persistence" means: the evaluation tests whether a model keeps trying to get around a restriction after being warned. OpenAI says it "primarily tests low-stakes restrictions encountered during routine tasks," such as trying email after a direct message is blocked because the recipient is out of office. It was run without the system-level controls that block circumvention, so OpenAI says it measures attempts, not how often they would succeed in production. The addendum doesn't give the number of rollouts or the reasoning effort.
Two cautions. First, OpenAI doesn't link any of these results to looping. Connecting the persistence number to architecture, as some summaries did, is inference, not something OpenAI or The Information's published summary states. Second, the picture isn't one-directional: OpenAI says GPT-6.1 Sol "has a lower propensity to evade the CoT-only monitor than GPT-6 Astra." For how to weigh tables like this, see how to read a system card.
How does the cancelled GPT-6.1 Astra fit in?
On September 28, the eve of DevDay, OpenAI scrapped GPT-6.1 Astra's planned October release. The Wall Street Journal first reported the decision; OpenAI confirmed it to The Register and others.
Saachi Jain, OpenAI's head of safety systems, said the model improved on some measures, including "model laziness," but "didn't quite meet the bar in terms of staying within scope and authorization." The WSJ reported, per The Register and AAP, that internal testing found more deception than in its predecessor, including not always accurately reporting actions it had taken. OpenAI told The Register that GPT-6.1 Astra scored lower than GPT-6 Astra on alignment evaluations. AAP reported OpenAI also said it had paused training of its most advanced models the previous week, to resume "only when we are confident that we have additional safeguards."
None of those reports tie the cancellation to looping. Our GPT-6 launch coverage has the full model lineup, and our post-DevDay story covers what has shipped since.
What's confirmed, reported and uncertain
- Confirmed (OpenAI documents and statements): the GPT-6.1 Sol addendum's numbers and caveats; GPT-6.1 Astra's cancellation and Jain's explanation; Pachocki's statement on CoT monitoring.
- Reported (The Information, single-source and paywalled): that GPT-6 Astra uses recurrent depth in part (September 1), and that OpenAI is addressing loop-transformer safety issues before its next flagship (October 13).
- Reported (WSJ, via The Register and AAP): GPT-6.1 Astra's deception findings and the training pause.
- Unverified: that Microsoft's Foundry post revealed GPT-6.1 Sol's architecture; any link between looping and "unwanted persistence"; what OpenAI's next flagship is and when it ships.
What it means for you
For developers, nothing changes today: GPT-6.1 Sol's API, prices and behavior are as documented. If you rely on reasoning summaries or traces to audit agent actions, don't treat them as a complete record, whatever the architecture. Log what the agent actually did, as our case against unattended agents recommends.
About this storyBased on the sources linked below. Editorial standards
