Anthropic tells the SEC it may be nearing its threshold for AI that accelerates AI

On November 2, 2026, Anthropic, PBC filed with the SEC an amendment to the IPO prospectus it submitted on October 26: a Form S-1/A with accession number 0001628280-26-068676 and file no. 333-299633. In it, the company writes that its internal estimates and expectations about future models, not current ones, indicate it could soon reach one of the thresholds in its Responsible Scaling Policy. The exact wording is that Anthropic “may imminently meet our Responsible Scaling Policy's definition of ‘dramatically accelerated progress attributable to AI’”. In the same document it adds: “We have not designated any of our current models as exceeding this threshold”. In other words, none of today's models has been classified as past the threshold.

A note on method that applies to the whole piece: the full text of the SEC filing was not read, and each sentence quoted here was confirmed through EDGAR's official full-text search.

To understand what this is about, you have to look at the RSP, the framework Anthropic uses to tie model capability levels to mandatory safeguards. One of its thresholds concerns AI that markedly speeds up research on AI itself, the scenario usually called recursive self-improvement. According to the official page, the version in force is 3.4, effective since July 8, 2026. The public page, however, still only carries the clarification introduced with version 3.1, not the updated wording of the threshold. That clarification describes it as the ability to compress two years of AI progress at the 2018–2024 pace into a single year. So this is about doubling the rate at which aggregate capabilities advance, not about the productivity of individual researchers.

On countermeasures, Anthropic writes that it is evaluating additional steps. These include updates to the RSP that would require human review and approval for certain decisions in its research and development processes, in the company's words “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic”. The company also says it does not consider it safe to let Claude build its successor models on its own without meaningful human supervision or understanding. Two warnings were already in the October 26 S-1. The first is “Using AI to help build more capable AI could also pose a considerable safety risk”. The second is the possibility that the company might only notice a significant acceleration in its own pace of development after it has happened. What is new, and appears only in the November 2 amendment, is the threshold that may be reached soon and the human review. The document does not say when the threshold might be crossed, by which future model, or with what concrete measures. The RSP revision is described as under consideration.

Where these sentences appear matters. An IPO prospectus is a formal statement to investors, subject to liability under SEC rules, not a blog post. It is also a risk-disclosure document, though, and some commentators read it as market communication too. The estimates are Anthropic's own, and no third party has verified them. According to TechCrunch and other outlets, the S-1/A still leaves the price range blank, and the planned ticker is ANT on the Nasdaq Global Select Market.

Recursive self-improvement usually turns up in essays and conference talks. Here it sits among the risk factors of a prospectus, next to a price range that is still empty. The company says the threshold could arrive soon, while the remedy it points to is a policy update still under consideration. If an RSP revision does come, I will read it with the care usually reserved for footnotes. — Olya