Anthropic filed an amended version of its IPO paperwork with the SEC on November 2, 2026, and one new paragraph caught my attention. The company wrote that its "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.'" In plain English, Anthropic thinks its next models may be close to speeding up AI research so much that its own safety rules treat it as a new kind of risk.
The same paragraph says Anthropic hasn't designated any current model as crossing that line, that it's still assessing the data, and that it will release further reports as its policy requires. None of this was in the original S-1 filed October 26. It showed up in the amendment a week later, along with the company's third-quarter results.
The line it's talking about comes from Anthropic's Responsible Scaling Policy, the rulebook that decides how careful the company has to be as its models get more capable. Version 3.4, in effect since July 8, 2026, describes the threshold as AI systems that can fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams of human researchers in fields where fast progress could threaten international security or shift the global balance of power. Its examples are "energy, robotics, weapons development, and AI itself." The policy's changelog says the threshold is meant to catch "the onset of dramatic recursive self-improvement," which is AI helping build better AI that then builds better AI still, faster each round. An earlier note described it as AI compressing two years of 2018-to-2024-style progress into a single year.
The filing is also blunt about why Anthropic might not see the moment coming. Most of its evidence about how fast things are moving comes from lagging indicators, and it says it "may in the future not identify, a significant acceleration in our rate of development until after it has occurred." It admits this has happened before.
Claude already does a big share of Anthropic's work
Anthropic calls this "closing the loop," where each version of Claude helps build the next one. The numbers in the filing show how far that's gone. The value of Anthropic's own internal Claude usage, priced at list rates, grew more than 250 times from January to September 2026. In April, that monthly value passed the company's monthly payroll costs, not counting stock compensation. Anthropic had about 5,300 employees at the end of September, and its trailing 12-month revenue per employee rose from $1.4 million to more than $6.6 million in a year.
The amendment adds two sentences that read like the company talking to itself. It says it doesn't expect to be able to safely use Claude "to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding," even if that becomes technically possible. And it says Anthropic is "considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic." The filing points to the precautions it has taken before, like delaying releases and limiting early access, as it did with Mythos through Project Glasswing, and warns that more of them could slow the company down against competitors that don't hold themselves back.
There's a reasonable skeptical read. Risk sections in an IPO filing are written to protect the company from lawsuits, so they tend to name every bad outcome a lawyer can imagine, and a warning about AI moving too fast also tells investors the product is getting very good. TechCrunch's Julie Bort noted that skeptics see it as partly marketing. Still, a company doesn't usually write "imminently" into an SEC filing by accident, and Anthropic has spent years arguing that this exact scenario is the one to worry about. We wrote about the human side of that in who checks the code AI writes.
The new third-quarter numbers
Revenue for the three months ended September 30, 2026 was $17.0 billion. For the first nine months of the year it was $33.2 billion, up from $2.5 billion in the same stretch of 2025, a 1,206% increase. Third-quarter revenue grew 1,127% from a year earlier, while compute and infrastructure costs grew 382%.
Anthropic still lost money on paper. Its third-quarter operating loss was $567 million, down from $3.1 billion a year earlier. On its own adjusted measure, which strips out stock compensation and some other items, it made $2.5 billion in operating income for a 15% margin. That margin was down from 19% in the second quarter, which the filing blames on $5.0 billion spent on short-term compute.
The price range and the number of shares are still blank in the amendment. Our rundown of the original filing covers the ticker, the share classes, and who controls the votes. For the self-improvement question, the next thing to watch is Anthropic's next Risk Report. The policy says it publishes one every three to six months, and the last was in August.
Sources
SEC: Anthropic, PBC Form S-1/A (Amendment No. 1) - Filed November 2, 2026. The imminent-threshold paragraph, the human-control and supervision statements, internal Claude usage, headcount, revenue per employee, and third-quarter and nine-month financials.
SEC: Anthropic, PBC Form S-1 - Filed October 26, 2026. The original filing, used to confirm which statements are new in the amendment.
Anthropic: Responsible Scaling Policy, Version 3.4 - Effective July 8, 2026. The automated R&D threshold, its examples, the recursive self-improvement changelog note, and the three-to-six-month Risk Report schedule.
Anthropic: Responsible Scaling Policy updates page - Version history, the earlier two-years-in-one framing, and the August 2026 Risk Report.
TechCrunch: Anthropic warns that it's very close to RSI capability - Julie Bort, November 2, 2026. Coverage of the amendment and the skeptical reading.
