Anthropic's amended IPO filing contains the clearest statement yet from a frontier AI company that its systems may soon speed up the work of building AI. The amended S-1, filed with the US Securities and Exchange Commission on 2 November, says internal estimates about future models indicate Anthropic "may imminently meet" its own threshold for "dramatically accelerated progress attributable to AI." TechCrunch reported the language the same day. We compared the passages on AI-accelerated research in the original S-1, filed on 26 October, with the same passages in the amendment. Three are new or rewritten, two in the risk factors and one in the business section, and reading them exactly matters more than any headline.
Change one: a warning moved into the past tense
Both filings warn that Anthropic's use of AI in its own research "may cause the capabilities of our models to advance more rapidly" than its safety work. Both say much of its evidence comes from "lagging indicators." The original then said: "we may not identify a significant acceleration in our rate of development until after it has occurred."
The amendment reads: "we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred." The new wording says this has already happened at least once. The filing does not say when, or how large the missed acceleration was.
Change two: the threshold sentence
The amendment then adds a new example: "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 continues: "We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data." It also says Anthropic is evaluating what meeting the AI research threshold would mean and will publish further reports, as its policy requires.
Two details are easy to lose. The claim is about future models, and it rests on Anthropic's own estimates rather than an outside measurement.
What the threshold means
Anthropic's Responsible Scaling Policy is the internal rulebook that ties new safety requirements to specific model capabilities. One set of thresholds covers AI research and development: the point at which AI meaningfully speeds up the building of AI itself, a scenario often called recursive self-improvement. The original S-1 already used that term once.
The policy's public change log explains how the bar has been drawn. In version 3.1, effective 2 April 2026, Anthropic clarified that language about AI that could "compress two years of 2018 to 2024 AI progress into a single year" means doubling the rate of progress in overall AI capabilities, rather than doubling the productivity of individual researchers. Version 3.4, effective 8 July 2026, revised the automated research threshold again. In a change log entry dated 10 February 2026, Anthropic wrote that confidently ruling out its AI R&D-4 threshold was "becoming increasingly difficult." The S-1 amendment is the first time the company has said publicly that crossing the line may be imminent.
Change three: humans stay in control
The third new passage appears in the business section, next to the existing line about recursive self-improvement. It opens with a confident line, saying Anthropic believes its safety investments will let it "responsibly realize substantially more leverage from Claude than we do today" in building AI models. It then writes that "we do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible." It adds that it is "considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development."
That is a statement of intent, and the word "considering" matters. No policy update has been published yet.
How to read it
A risk factor in an S-1 is written by lawyers to disclose what could go wrong, so companies lean towards including more risks. TechCrunch noted that skeptics might see an S-1 as "both disclosure and marketing," and that a claim to be close to self-building AI could fall into the second category. Both readings can be true. We think the most important change is the past-tense line, because it is an admission about the past, which is harder to dismiss as positioning than a forecast about the future.
For readers who use Claude at work, nothing changes this week. The filing describes Anthropic's research process, and the current models you use have not been designated as crossing the threshold.
What to do this week
- Read the three quoted passages before sharing a headline about them, since most summaries drop the words "future models."
- Watch for the Responsible Scaling Policy update and the further reports Anthropic says it will publish.
- If you brief leaders on AI risk, use the filing's own wording, and label it as Anthropic's estimate.
For what the same amended filing shows about Anthropic's own use of Claude, read our piece on Claude use passing Anthropic's payroll, and for the original filing, see what Anthropic's S-1 reveals.