Anthropic Tells Investors Its Own AI Line May Be Near

On November 2, Anthropic filed the first amendment to its IPO registration statement with the Securities and Exchange Commission. Among its risk factors is a sentence that does not appear in the S-1 the company filed on October 26 or in the October drafts before it, according to a search of the SEC’s full-text database. The company’s “internal estimates and our expectations regarding future models, rather than our current models,” it wrote, “indicate that we may imminently meet” the definition its Responsible Scaling Policy gives for “dramatically accelerated progress attributable to AI.” The same passage adds, “We have not designated any of our current models as exceeding this threshold,” and promises to “release further reports.”

The line is one Anthropic drew for itself. Under version 3.4 of the policy, in effect since July 8, the company treats the threshold as met if it determines either that its models could fully substitute for its research scientists and engineers at competitive cost, or that there is “dramatic acceleration” of AI progress for reasons likely tied to automating research. The second path counts when Anthropic observes or expects the rate of progress in aggregate AI capabilities to double—“as much progress in one year as one would see in two years at baseline”—against both the rate it would otherwise expect and the fastest sustained rate it has seen without significant help from AI, and judges the doubling plausibly due to automation rather than to more staff, compute or general productivity, such that continuing the trend would likely bring even greater acceleration. The policy’s own changelog concedes that this threshold “has proven difficult to operationalize.”

Under that definition, an expectation can be enough to meet the bar. The November 2 sentence is still a forecast, not a designation, and the filing keeps that distinction. Readers should too.

What gives the forecast its weight is an older passage that has just been rewritten. Since at least a draft dated October 1, Anthropic’s registration filings have said that much of its evidence on the pace of AI-driven acceleration is “derived from lagging indicators,” and that it may not identify a significant acceleration in its rate of development until after it has occurred. In October that was a possibility. The November amendment adds a record: the company says it has, in the past, not identified such an acceleration until after it had happened. The filing does not say when that was, how late the recognition came, or whether that episode would have met the policy’s threshold.

None of this means Anthropic lacks a process. The policy commits it to publish a Risk Report every three to six months, on a calendar. It also requires a separate analysis “within 30 days of determining” that an internally deployed model poses automated-research or misalignment risks significantly beyond anything it has analyzed publicly. The company says it is working toward public external review of its Risk Reports and seeks outside feedback on them; its Long-Term Benefit Trust can request a review, and reviewers must be approved by the trust and free of financial ties to Anthropic. When a full review takes place, reviewers receive the report within a week of its submission to the board and the trust and are asked for public commentary within 30 days.

The calendar keeps running whatever happens. The clocks that answer to this particular line do not. The 30-day analysis starts from the company’s determination. So does the policy’s minimum guarantee of a full external review: it applies when a report covers a “highly capable” model and is significantly redacted, and a model counts as highly capable “if we conclude that it crosses the threshold for automated AI R&D.” Put that beside a filing which says the company’s evidence of acceleration comes in late, and has come in late before, and the delay in noticing is added before those deadlines start to run.

Outside review has already pointed at the instruments. In May, the evaluation group METR reviewed the automated-research section of Anthropic’s February Risk Report, using non-public material the company shared. METR agreed that the risk of catastrophe from Claude Opus 4.6, or less capable Anthropic models, automating research was very low, but wrote: “we think the evidence presented in the report is inadequate to establish this.” It urged the company to report more leading indicators of AI progress. Six months later, the investor document still describes evidence that lags.

The public evidence Anthropic itself has offered shows why measurement is hard. In a June essay, “When AI builds itself,” two of its staff wrote that “in the second quarter of 2026, the typical engineer was merging 8× as much code per day as they were in 2024.” That counts code merged per engineer. It can mean faster progress, or simply more code. The same essay said, “We are not there yet.” Five months on, the company’s word is “imminently.”

Anthropic can point to how much of this it chose to put on the record. It wrote the threshold down and published it, along with a forecast that it may meet it and an admission that its evidence has trailed events. It has invited outside reviewers under conflict-of-interest rules. Elsewhere in the same risk factors, it says it is considering policy updates that would require “human review and approval” of certain decisions in its research and development processes—an option under consideration, with no date attached. In June it said it would slow down or temporarily pause if other frontier developers did so “in a verifiable manner.” Admitting that a gauge reads late is a kind of accountability in itself: a company that says so is easier to hold to account than one that claims its readings are instant.

The policy’s history follows the same pattern. In May 2025, launching Claude Opus 4, Anthropic said in a blog post that it had activated stricter ASL-3 safeguards while it had “not yet determined whether Claude Opus 4 has definitively passed the Capabilities Threshold.” It chose to act before it was sure, and the call was its own. Acting early is the policy’s strength. It also means that whatever a model can actually do, the official designation is the company’s judgment.

Other dangerous industries count from a different moment. A U.S. nuclear plant operator must notify the Nuclear Regulatory Commission within one, four or eight hours of specified events, under categories written by the regulator. Those clocks run from the event. The policy’s 30-day clock runs from the company’s own determination.

The forecast’s first audience says something too. The sentence about future models appears as a risk factor in an offering document, addressed to investors weighing risks to their money. Anyone can read it there. A risk factor settles whether investors were told. It leaves open when people outside the company should have a say.

The further reports Anthropic promises will show whether that changes. Two things would matter most: leading indicators of the kind METR asked for, published ahead of events, and at least one clock tied to this line that starts from something other than the company’s own determination. Until then, the judgment that the line has been crossed stays inside the company, where the chief executive and the Responsible Scaling Officer give Risk Reports final approval and, in some cases, the board and the trust must approve as well—and by the company’s own account, the evidence that reaches them can arrive after the fact.