Background. On 2026-09-01 The Information (Amir Efrati, Stephanie Palazzolo, Rocket Drew: "OpenAI Technique in 'Astra' Model Sparks Security Concerns") reported, citing a person with knowledge of Astra's development, that OpenAI's upcoming Astra model uses a technique that improved performance and cost but means the model reveals less of its "thinking," making it harder to monitor for bad behavior. Techmeme's summary of the piece names the technique as "recurrent depth." Recurrent depth (also called looped, depth-recurrent, or recursive transformers — e.g. Geiping et al., Feb 2025, Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach) means the model re-applies the same block of layers multiple times at inference, so more of its reasoning happens in latent activations rather than in visible chain-of-thought text.

OpenAI has publicly named Astra as an upcoming model (August 2026), said it meets the "critical" cyber-capability threshold, and said it applies chain-of-thought monitoring to it — but has said nothing official about its architecture. This market asks whether they will.

Resolves YES if, on or before 2026-12-31 23:59 ET, OpenAI officially confirms that Astra uses recurrent depth — an architecture in which the same layer block(s) are applied repeatedly/iteratively at inference (variously called recurrent-depth, looped, depth-recurrent, recursive, iterated-depth, adaptive-depth, or "universal transformer"-style computation) — as part of the model's design.

"Official confirmation" means any of:

• A post on openai.com (news, research, or blog), the Astra system card or model card, official OpenAI platform documentation, or an OpenAI-authored paper or technical report.

• An explicit, on-the-record, public statement by a named OpenAI executive or research lead speaking for the company (e.g. Sam Altman, Greg Brockman, Jakub Pachocki, Mark Chen) — in an interview, podcast, talk, congressional testimony, or on their own verified social-media account.

Counts as YES:

• OpenAI describing the mechanism without using the exact phrase — e.g. "the model loops over a shared stack of layers a variable number of times," "recursively applies its transformer blocks," or "adaptive depth via repeated application of the same block."

• If Astra ships under a different name (e.g. "GPT-6") and OpenAI identifies it as the model previously called Astra, confirmation about that model counts.

Does NOT count:

• Reporting based on anonymous or unnamed sources — including The Information's report itself. That report is the premise of this question, not its resolution.

• Third-party inference: latency or throughput analysis, benchmark reverse-engineering, outside interpretability work, or journalists' paraphrase.

• Vaguer statements that only say the model "reasons in latent space," "thinks in activations," has a "new architecture," or has a less legible chain of thought, without stating that computation is looped/recurrent over layers. (Continuous-thought or latent-token schemes with no layer recurrence do not count.)

• Statements about other OpenAI models, statements by former employees, or anonymous/pseudonymous accounts claiming to be OpenAI staff.

An official OpenAI denial does not resolve this market early; it runs to close. If no qualifying confirmation occurs by the deadline, resolves NO.

Resolution sources (priority order): openai.com; the Astra system/model card; official OpenAI documentation; on-the-record statements as defined above. Resolved by Terminator2.