Anthropic told the SEC it may soon cross its own AI R&D line. The evidence is mostly its own.

On November 2, Anthropic filed its first amendment to the IPO registration statement it made public on October 26. One risk factor gained a new sentence. In the amended S-1, the company writes that “our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI.’” It adds that it has not designated any current model as exceeding the threshold and that it “will release further reports in the future.” The reporting is accurate; we read the filing on EDGAR. The thesis of this piece is simple. When a lab tells securities regulators it may soon cross its own recursive self-improvement line, the question is no longer what it believes. It is who can check, with what data.

Two other edits matter. The original S-1 warned that, because its evidence comes from “lagging indicators,” Anthropic “may not identify” a significant acceleration until after it has occurred. The amendment now reads “we have in the past not identified, and may in the future not identify” one. That is an admission that it has already happened at least once. The amendment also adds that Anthropic is “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development,” including “human review and approval of certain decisions” in research as models do more of it. Considering is not adopting.

What is the line? Anthropic's Responsible Scaling Policy v3.4, in force since July 8, treats its automated R&D threshold as met if either “(1) our models would be able to fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs (i.e., within a factor of 5); or (2) there is ‘dramatic acceleration’ of the pace of AI progress for reasons that likely relate to the automation of AI R&D.” The second test requires that “we observe or expect double the rate of progress in AI aggregate capabilities” relative to the fastest rate seen over at least three model generations without major AI help, and that the doubling plausibly comes from automated research. A footnote defines doubling as “as much progress in one year as one would see in two years at baseline.” Two details stand out. The phrase the filing puts in quotation marks does not appear verbatim in the published policy. And the word “expect” means a forecast alone can satisfy test (2). By the policy's own text, the filing's “expectations regarding future models” are not a hedge. They are the relevant evidence. The policy's own changelog says this threshold “has proven difficult to operationalize,” and it has been revised twice since April.

What happens when it is crossed is weaker than many readers will assume. Version 3 of the policy splits Anthropic's own plans from “recommendations for industry-wide safety” and says plainly that “we cannot unilaterally and unconditionally commit” to the latter. Those recommendations, for this threshold, include security around RAND SL4 and controls that bind insiders “up to and including the company’s CEO.” The company's own plan is logging “all critical AI-development activities,” alignment assessments, and Risk Reports. The hard triggers are narrow. A model over the line counts as “highly capable,” which requires a full external review only if the Risk Report is “significantly redacted.” A commitment to delay development applies if Anthropic has, or “will imminently develop,” such a model and has “clear evidence that no other competitor will soon develop” one. In a race, that condition is unlikely ever to be established.

The other labs draw the line elsewhere. OpenAI's Preparedness Framework (version 2, April 2025) sets High self-improvement at the equivalent of giving every researcher “a highly performant mid-career research engineer assistant,” which requires High security. Critical means “a superhuman research-scientist agent” or a generational model improvement “in 1/5th the wall-clock time of equivalent progress in 2024 … sustainably for several months,” and it carries a commitment to halt further development until safeguards are specified. Google DeepMind's Frontier Safety Framework 3.1 defines its first acceleration level as AI having been used so that progress is “substantially accelerating from historical rates,” with no number, and recommends Security Level 3. So Anthropic's 2x is an early warning, OpenAI's 5x is a halt, and DeepMind's is a judgment call. DeepMind also says it may assess these thresholds using “sources of information about our own progress.” All three frameworks are, at this level, self-measured.

The best view of the data comes from the two most recent documents. Anthropic's August 2026 Risk Report says Claude “now authors a large majority of the code merged into our production codebases” and that internal R&D is significantly faster with AI, “but not yet by a factor of 2.” It rated the risk low but said it was less confident than before because its task-based evaluations have “saturated.” Its leading indicators “are sensitive and not included in the public version.” On November 5, METR, which read the unredacted version, published a review. It agreed catastrophic risk was low but found it “plausible” the 2x threshold had already been crossed by mid-July. It put the acceleration implied by Anthropic's private estimate at about 1.97x, gave its own range as 1.3x to 2.8x, and concluded that “the public report overall understates the level of acceleration that was occurring.” It also noted the review's scope was “mutually negotiated” and that Anthropic approved the redactions on October 29, four days before the amended filing.

Independent measurement cannot settle the question either, for a reason worth stating. METR's time-horizon benchmark shows the length of software tasks agents complete half the time doubling roughly every 129 days since 2023. Its highest point, an early Claude Mythos Preview from April at 17.4 hours, has a 95% interval of about 8.5 to 55 hours. METR itself says results above 16 hours are unreliable. RE-Bench, built for AI R&D tasks, found 2024 agents scored 4x higher than human experts with two hours each but half as well with 32. Productivity studies are harder still. In METR's 2025 randomized trial, developers expected AI to make them 24% faster, believed afterward it had made them 20% faster, and were measured 19% slower. The 2026 follow-up found an 18% speedup with a confidence interval spanning zero, and METR said the design broke because developers declined to work without AI. The instruments are saturated, and the perception gap runs one way. Neither pattern favors taking self-reported acceleration on trust.

Here is what should follow from a filing like this. First, the threshold tests should be defined by metrics that are published in advance, even if the values are redacted. Today the decisive leading indicators are secret in definition as well as value. Second, external review of an approaching threshold should have a fixed scope and data access set before the work starts, not negotiated during it. Third, the industry needs one shared acceleration metric. A 2x early warning, a 5x halt and an unquantified “substantially” cannot be compared, and a regulator cannot enforce a race among self-defined lines. Fourth, the human-control rules Anthropic says it is considering should be written into the policy before the threshold is met, not after. Anthropic deserves credit for saying this in a document it can be held to; no other lab has said as much. But a disclosure is not a measurement. Right now the clearest evidence that the line is close comes from the company's own estimate and from the one outside reviewer it allowed to read it.