GPT-6 Astra's per-forward-pass computational depth is within a factor of two of GPT-4's.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The claim rests on a single public statement by OpenAI's chief scientist, Jakub Pachocki, posted on 2 September 2026 after The Information reported that GPT-6 Astra uses recurrent (looped) depth: "The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4." Pachocki did not deny that Astra uses recurrence, and the reporting itself said the recurrence was deliberately limited, which fits his figure.
The statement comes from the best-placed possible source, is specific, and would be costly to have gotten wrong, so it is good evidence. Its limits are definitional and evidential. Neither model's depth has been disclosed: GPT-4's layer count is known only from unofficial reports. "Depth of the computation graph" is not defined, and for a model whose loop count may adapt per token it is unclear whether the figure refers to typical or maximum depth. No independent measurement exists or is expected. Commentators have noted that the figure is consistent both with a small loop count and with no looping at all. The claim is best read as credible developer testimony that has not been verified.
Full reasoning: the evidence and decisions behind this verdict
Prior model: the claim cannot be verified from outside because both architectures are proprietary, so it turns on how reliable the developer's testimony is. A well-placed, specific public statement by a chief scientist on a technical fact is fairly reliable. The remaining risk is mostly ambiguity (unique versus unrolled layers, typical versus maximum recurrence) rather than falsehood.
Evidence found: the primary statement is Pachocki on X, 2026-09-02, x.com/merettm/status/2095023204993490967. Secondary pieces quote it consistently: a Medium architecture analysis (medium.com/@lyx_62906/the-gpt-6-astra-architecture-puzzle-05fd2ff781af), which calls it a direct public statement and uses it as a consistency check; a Substack piece (theinterestingparts.substack.com/p/the-caveat-is-about-the-caveat-reading), which notes that it is equally consistent with a small loop count and with no loop; and a blog piece on looped transformers (saulius.io/blog/what-are-looped-transformers-recurrent-depth). None of these offers independent evidence, and none disputes the figure. The Information's report that recurrence was deliberately limited fits it.
The claim's wording, "per-forward-pass computational depth", matches Pachocki's "depth of the computation graph" closely enough to count as the same proposition. The claim is atomic: it rests on testimony, not on a decomposition into reusable propositions, so no subclaims were added. Credence is 0.8: discounted from the testimony for the undefined metric and possible adaptive-depth edge cases, not for suspected deception. The status would move with a technical disclosure (a system card architecture section or a paper) or with credible contrary reporting.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4.
OpenAI's chief scientist responding to reporting that GPT-6 Astra uses recurrent depth, saying he wanted to prevent a race into unmonitorability kicked off by confused reporting.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.