A frontier model reasons in a form people cannot read

Where things stand (as of 12 Oct 2026)

The technique is probably already in use, but no developer has said so. The Information reported on 1 Sep 2026 that OpenAI's GPT-6 Astra uses "recurrent depth", also called a looped transformer: the same block of layers is run several times on a token without writing anything out. Independent no-reasoning tests (LessWrong, 30 Sep) and a deleted Microsoft page ("two inference passes instead of three") suggest GPT-6.1 Sol uses it too. None of that counts. Only developer-authored material dated on or after 16 Sep counts.

What OpenAI has actually written:

• Astra system card (published 3 Sep; updated 9, 22 and 29 Sep). I downloaded it and searched the text. It has no words about recurrence, loops, latent states or neuralese. It reports lower monitorability and an order-of-magnitude jump in no-reasoning ability. It says OpenAI is "quite confident" that the controllability change is not "differentially due to any architectural changes", and that it will "share more about our approach in the coming months."
• GPT-6.1 Sol addendum (29 Sep): no architecture content at all.
• Pachocki's X post: says computation depth is "within a factor of two of GPT-4". It doesn't count (employee social media), and it frames the change as depth, not hidden reasoning.
• Secrecy: OpenAI fired three safety researchers in early October amid a leak probe. They deny being the source of the architecture leak. This suggests OpenAI treats the architecture as sensitive.

Other covered developers:

• Zhipu. Its interim report (Sep 2026) says a Loop Transformer, which reuses layers to "think more deeply", has been validated only in preliminary experiments. Tang Jie put it on the next-generation roadmap. Zhipu frames it as "effective depth" and stresses "reviewable traces".
• Google DeepMind. Its essay "The case for reasoning transparency" (16 Sep) argues for keeping chain-of-thought architectures and limiting how much reasoning can happen out of sight. Gemini 4 Argon (30 Sep) relies on monitoring its written reasoning.
• Anthropic. The Opus 5.5 announcement (22 Sep) says only that it hopes interpretability tools will reduce its reliance on auditing chain of thought.

I found no qualifying publication so far.

Reference class and base rate

• Track record: no covered developer has ever made the required statement. That includes OpenAI with Astra, where the mechanism reportedly exists.
• How papers describe looped models: I searched the full text of Ouro (ByteDance Seed) and Loopie, the largest open looped models. Neither contains "human-readable", "interpretable" or "legible" language. Ouro instead claims its latent reasoning is more faithful than chain of thought.
• What that implies: developers who adopt loops tend to call them depth or efficiency, not unreadable reasoning.
• Conditional estimates: given adoption and a description, I put the chance of a qualifying unreadability statement at about 15–30% for Chinese open-weight labs. For Western labs it is about 50–70% once they describe the mechanism at all, but they rarely disclose architecture.

Paths to YES

1. OpenAI confirms the looped design in a system card, monitorability paper or industry disclosure standard, and says the looped computation is not visible in the chain of thought. Later models with more recurrence would make this more likely.
2. Zhipu, DeepSeek, Qwen, Kimi or Seed ship a looped or latent-reasoning flagship with a tech report that says the latent steps have no human-readable chain of thought. Zhipu's next generation, likely in 2027, is the clearest candidate.
3. Anthropic or Meta adopt latent reasoning. Anthropic would be candid, especially if it relies on activation-translation tools; the criteria still count a model whose developer offers a translation.
4. A broader shift by 2028–31 driven by efficiency (OpenAI's CFO cited 68% fewer output tokens), automated research, and the fact that only one statement from any of 10 developers is needed.

Paths to NO

• Architecture secrecy at closed labs.
• Strong monitorability norms: DeepMind, Anthropic and OpenAI have all endorsed an industry-wide commitment, and limits on opaque serial depth may come.
• Chinese labs framing loops as depth.
• Loop counts staying small (2–3).
• A strict resolver reading of "explicitly states" and "intermediate reasoning".

Forecast

• Hazard rates: about 2.5% per quarter near term, rising to about 2.8% in 2027–28 as next-generation models land (Zhipu's loops, possible OpenAI disclosures), then easing to about 2.2% by 2031.
• Cumulative: about 2.5% by end-2026, 12% by end-2027, and 42% by end-2031. I computed the hazard series in Python.
• Strongest case against this estimate: the requirement has never been met, even though the mechanism exists. That pushes toward about 20%. The long horizon and pressure from many labs push toward about 60%.

Sources

• OpenAI, Safety overview: GPT-6 Astra: openai.com/index/safety-overview-gpt-6-astra (3 Sep 2026)
• GPT-6 Astra system card: deploymentsafety.openai.com/gpt-6-astra (updated 29 Sep)
• GPT-6.1 Sol addendum: deploymentsafety.openai.com/gpt-6-1-sol (29 Sep)
• Pachocki, An Alien Mind: openai.com/index/an-alien-mind (6 Sep)
• Fortune (3 Sep) and TechCrunch (2 Sep) on The Information's report
• Pachocki on X: x.com/merettm/status/2095023204993490967
• LessWrong posts by Rauno Arike (2 Sep and 30 Sep)
• Pillitteri on the Microsoft page (Oct 2026)
• TechCrunch on the researcher firings (8 Oct)
• DeepMind Institute essay: institute.deepmind.com/essays/the-case-for-reasoning-transparency (16 Sep)
• Gemini 4 Argon blog (30 Sep)
• Anthropic, Claude Opus 5.5 (22 Sep)
• Zhipu 2026 interim report: stockn.xueqiu.com/02513/20260924938161.pdf
• Locsic on Zhipu's Loop Transformer
• Ouro: arXiv 2510.25741; Loopie: arXiv 2607.16051
• Redwood Research, "Latent reasoning architectures would likely undermine CoT" (23 Sep)

Question. By the last day of each quarter, has a covered developer released or described a frontier model whose intermediate reasoning is carried in a form other than text, which the developer says people cannot read?

Definitions.

• Frontier model: an actual trained model that its developer identifies as a frontier model, or as one of its most capable general-purpose or reasoning models at the time of the qualifying description. Internal and unreleased models count; a proposed architecture or small-scale proof of concept alone does not.
• Intermediate reasoning: inference-time processing of intermediate problem-solving states before producing an answer or action. The developer must describe a mechanism that carries reasoning states forward without expressing those states as text, such as continuous latent thoughts, recurrent hidden-state reasoning, or learned nonlinguistic codes. Merely having neural activations, embeddings, attention states, or a key-value cache does not establish this condition.
• Other than text: the reasoning states are not sequences of ordinary text tokens representing natural language, code, mathematical notation, or another human-readable symbolic language. Serializing latent vectors as numbers does not make them text reasoning. Hybrid systems with both textual and nontextual intermediate reasoning count; no minimum fraction of nontextual reasoning is required.
• People cannot read: the developer explicitly states that the nontextual reasoning representation is not human-readable, cannot be directly interpreted by people, or has no human-readable chain of thought for those reasoning steps. This does not require a claim that interpretation will forever be impossible. A generated explanation, translation, or summary does not disqualify the model if the developer says the underlying representation is not directly human-readable.

Resolves YES if qualifying developer-authored material establishes all of the following for the same model:

1. The model exists and meets the frontier-model definition.
2. It actually uses the nontextual intermediate-reasoning mechanism, rather than merely being capable of an untested or proposed modification.
3. The developer makes the human-unreadability statement defined above about that mechanism.

Public access to the model, publication of weights, and external verification of its reasoning are not required. Multiple qualifying documents may jointly establish the conditions; resolve on the publication date of the last necessary document.

Exclusions (resolve NO).

• Textual chains of thought that are withheld, hidden in a user interface, encrypted, compressed through an ordinary reversible text encoding, or available only to the developer.
• Claims only that reasoning is difficult to monitor, unfaithful, incomplete, opaque in a general sense, or not fully understood, without the specified statement about the nontextual representation's human unreadability.
• Nontextual inputs or outputs, including images, audio, and video, without a qualifying intermediate-reasoning mechanism.
• Research demonstrations not identified as frontier models, including the small-scale Coconut demonstrations in Training Large Language Models to Reason in a Continuous Latent Space (arXiv:2412.06769); an actual frontier implementation must independently satisfy the conditions.
• Rumors, third-party architectural inferences, and announcements of plans to build such a model.

Resolving source. Developer-authored system cards, technical reports, research papers, model documentation, or official statements on the covered developer's own domain, under Public confirmation (A2), narrowed here to developer-authored sources because the question requires what the developer says. Third-party reports cannot substitute for that statement. Apply the Resolution window and Resolution date (A3) conventions; documents used to establish the qualifying conditions must fall within the Resolution window.