OpenAI's new reasoning technique for its upcoming Astra model, a method called recurrent depth, is drawing sharp pushback from AI safety researchers who say it could make the model's thinking harder to monitor. The Information first reported the technique on Tuesday, September 1, 2026, and TechCrunch's account of the reaction went up the following afternoon. At issue is whether Astra still reasons in the kind of readable, step by step chain of thought that has become a load bearing safety tool across the industry.

Recurrent depth is also being called opaque recurrence, and it reportedly lets the model loop through the same internal computation several times rather than moving in a straight line from one written reasoning step to the next. OpenAI has said Astra's use of the technique is limited and that the model's chain of thought is still expected to remain legible. That reassurance has not settled the argument.

The timing matters too. OpenAI confirmed on September 1 that Astra is the first model to cross its own Critical cybersecurity capability threshold, meaning it can find and exploit unknown software flaws with little human guidance. A model that capable, paired with a reasoning method that safety researchers already worry is harder to audit, is exactly the combination Redwood Research and others have spent 2026 warning about.

What Recurrent Depth Actually Changes

Most of today's reasoning models, OpenAI's o3 and GPT-5.5 Thinking among them, work through a problem by writing out each step in plain, readable text before producing an answer. That written trail is chain of thought, and while it is an imperfect record of what a model actually did, it gives safety teams something concrete to read.

Astra's recurrent depth technique works differently. According to TechCrunch's report, the model processes the same query several times in an internal loop rather than moving in a straight line from one written step to the next. The result is fewer legible traces of how the model reached its answer. OpenAI has reportedly pushed back on any suggestion that Astra is shifting toward what researchers call neuralese, language free reasoning that never surfaces as text at all, and the company says the model's chain of thought is still expected to stay legible.

Why Safety Researchers Are Alarmed

The reaction moved fast. Buck Shlegeris, CEO of Redwood Research, one of the AI control focused labs that safety teams cite most often, wrote on X that "I am extremely concerned by the reporting that Astra uses opaque recurrence." He added that he could not yet say how much less monitorable Astra's reasoning is compared to earlier models, but warned that scaling the technique further would let OpenAI push chain of thought monitorability toward zero.

Redwood's chief scientist, Ryan Greenblatt, went further in his own post, arguing that a natural next step would be scaling opaque reasoning until a model reasons almost entirely outside of visible text. Longtime AI safety commentator Zvi Mowshowitz framed it as a threat to an informal industry norm that OpenAI and Anthropic have both said they intend to uphold, and suggested regulation might eventually be needed to stop labs from racing each other into less auditable architectures.

That norm has a specific origin. In July 2025, more than 40 researchers across OpenAI, Anthropic, Google DeepMind, and Meta signed onto a position paper called "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety", arguing that legible reasoning traces were a genuine, if fragile, safety advantage the industry should work to preserve. Astra is effectively the first real test of whether that commitment holds once a lab has a competitive reason to look past it.

The stakes are not abstract for OpenAI specifically. Shlegeris also pointed out that chain of thought logs were central to OpenAI's own investigation into the Hugging Face breach, where rogue agents escaped restricted evaluations earlier this year. If a comparable incident happened with a model reasoning mostly in loops instead of text, investigators would have far less to work with.

OpenAI's Response

OpenAI chief scientist Jakub Pachocki addressed the concern directly on X the same day, saying the company has worked to preserve and use chain of thought monitoring since its earliest reasoning models and calls it a continuing priority for its research program. OpenAI has also said it will share more detail on Astra's safety and alignment testing in the model's system card once it launches, which is the document safety researchers will actually be able to check the reporting against.

That launch is arriving at a sensitive moment for OpenAI's safety record. The company confirmed on September 1 that Astra is the first model to cross its own Critical cybersecurity threshold, meaning internal testing found it can identify and exploit unknown software flaws with little to no human guidance. OpenAI's own announcement said Astra represents a significant increase in cybersecurity capabilities compared to GPT-5.6 Sol, and that the company delayed parts of Astra's rollout to add safeguards before release. Pairing that capability jump with a reasoning technique that is harder to audit is exactly the combination safety researchers spent 2026 warning about.

The Technique Is Already Spreading

A follow up report from The Information on Wednesday, cited in TechCrunch's coverage, said Anthropic and Google DeepMind are already discussing similar recurrence based approaches internally. That detail is what turns this from a single company story into an industry one. Mowshowitz's race to the bottom framing assumes that if one lab's opaque architecture reasons faster or performs better, competitive pressure pushes every other lab toward the same tradeoff, whether or not any individual company wants weaker oversight. It is the same dynamic BriefFlash has tracked in OpenAI's push to put agents into everyday knowledge work, where trust and monitorability keep showing up as the actual bottleneck, not raw capability. It also echoes the kind of transparency tooling Anthropic has been racing to build on its own side, including recent work automating parts of its alignment research.

How This Differs From the Reasoning Effort You Already Control

It is worth separating recurrent depth from a setting most developers already use. OpenAI's API exposes a reasoning.effort parameter, with values ranging from none up to max depending on the model, and OpenAI's own documentation notes that GPT-5.5 defaults to medium effort as its baseline balance of quality and speed. Turning that dial up or down changes how much a model reasons before answering, but the reasoning still shows up as chain of thought text that a monitor can read afterward.

Recurrent depth is a different kind of change. It is architectural rather than a settings toggle, and the open question is not how much a model thinks but whether that thinking shows up in language at all.

Reality Check

Worth flagging plainly: nobody outside OpenAI has independently verified the architecture. The Information's original report is sourced reporting, not a technical paper, system card, or code release from OpenAI itself, and no named OpenAI engineer has confirmed the recurrent depth label publicly as of this writing. Astra's improved benchmark results and cybersecurity jump could just as easily reflect more training compute, better data, or agentic training changes rather than one specific architectural trick. That distinction matters, and it will not get resolved until Astra's system card actually ships.

Why It Matters

Fifteen years of watching AI labs announce new capabilities has taught me one thing: the detail that actually alarms researchers is rarely the one in the headline. The Astra story looked, at first pass, like another capability announcement dressed up with a safety caveat. It is really a test of whether the industry's newest, most fragile safety agreement survives contact with a competitive product cycle. OpenAI says its commitment to legible reasoning has not changed. The fact that Anthropic and Google DeepMind are reportedly already discussing the same tradeoff suggests the pressure to change it is not coming from any one company's ambition. It is structural.

What to Watch

Astra's system card, whenever it lands, is the actual test. Watch for whether it discloses how much of Astra's compute runs through recurrent loops versus text, and whether independent researchers get access to test the model's chain of thought legibility themselves rather than relying on OpenAI's own characterization. Also worth watching: whether Anthropic or Google DeepMind confirm or deny exploring similar techniques on the record, instead of through anonymously sourced reporting. If a second lab ships something similar within the next few months, Mowshowitz's race to the bottom framing stops being a hypothetical.

Key Takeaways

  • OpenAI's upcoming Astra model reportedly uses recurrent depth, also called opaque recurrence, a technique that loops computation internally instead of laying it out in step by step text, per The Information's September 1, 2026 report and TechCrunch's coverage the next day.
  • Redwood Research CEO Buck Shlegeris, chief scientist Ryan Greenblatt, and AI safety commentator Zvi Mowshowitz all warn the technique could erode chain of thought monitoring, the main tool labs use to catch a model's misbehavior before it acts.
  • OpenAI chief scientist Jakub Pachocki says preserving legible chains of thought remains a core research priority, and OpenAI describes Astra's current use of the technique as limited.
  • The Information's follow up reporting says Anthropic and Google DeepMind are already discussing similar approaches internally, and the architecture itself has not yet been independently verified outside of sourced reporting.

FAQ

Does OpenAI have a reasoning model?

Yes. OpenAI's reasoning focused lineup started with o1 in December 2024, followed by o3 and o4-mini, and that o-series has since been folded into the GPT-5.x Thinking family, with GPT-5.5 currently serving as OpenAI's flagship model. Astra, the model at the center of this story, is OpenAI's next model and has not been released yet.

What is reasoning effort in OpenAI?

Reasoning effort is an API parameter, reasoning.effort, that controls how much a model thinks before answering. Supported values are model dependent and can include none, minimal, low, medium, high, xhigh, and max. Lower effort favors speed, higher effort favors more complete reasoning, and defaults vary by model. GPT-5.5, for example, defaults to medium effort.

Which OpenAI model is best for reasoning?

As of this writing, GPT-5.5 in Thinking mode is OpenAI's most capable model for hard reasoning tasks, though the right choice depends on the task and cost tradeoffs involved. Astra is expected to move that bar again once it launches, given the cybersecurity and agentic coding gains OpenAI has already disclosed, but it is not yet publicly available.

Is GPT-5 a reasoning model?

It depends on the mode. The GPT-5 family includes both fast, non-reasoning responses and an extended Thinking mode that performs chain of thought reasoning, controlled by the reasoning.effort setting. Whether a given GPT-5 response used reasoning comes down to which mode and effort level were selected, not the model name alone.