Buried in the paperwork for one of the biggest tech IPOs ever is a sentence that deserves your full attention: in September 2026, more than 90% of the contributions to Anthropic's own codebase were written by Claude — the company's own AI. And in the same filing, Anthropic warns that AI which improves itself may be much closer than most people think.
What the filing actually says
The details come from Anthropic's amended S-1 registration statement (an S-1/A), filed with the US Securities and Exchange Commission on 2 November ahead of its planned stock market listing, and first reported by TechCrunch. Two disclosures stand out:
- Claude wrote over 90% of the code contributions to Anthropic's codebase in September 2026.
- Engineers' daily code output in the third quarter was more than 45 times its 2024 level — because each engineer now works alongside Claude.
Then comes the warning. In the filing's own words, the company's internal estimates indicate it may "imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI.'"
What is recursive self-improvement (RSI)?
That dry sentence describes a milestone researchers call recursive self-improvement: the point where AI gets good enough at building AI that progress starts feeding on itself. Under Anthropic's own safety policy, the threshold means AI systems either replace the company's entire research workforce at competitive cost, or double the overall rate of AI progress through automated research and engineering. It does not mean a conscious machine building itself in secret — it means the speed of AI development itself doubling because AI is doing the developing.
The other half of the story
Read the qualifiers before you panic. Anthropic states clearly that no current model has crossed that threshold — the warning is about what its internal projections expect from future models. The company also says it does not expect to be able to safely let Claude design and build successor models fully on its own, without meaningful human supervision, even if that becomes technically possible. And remember where this is written: an IPO filing's risk section is a legal document, where companies are obliged to state their worst-case risks plainly, however unlikely.
Why you should care anyway
Even the cautious version is remarkable. The tool writing 90% of a top AI lab's code is no longer a demo — it is the workforce. Every doubling in how fast these labs can work shortens the gap between today's AI and whatever comes next, and it is exactly why Anthropic is telling investors, in writing, that the acceleration itself has become a risk factor. When the company building the AI says the AI may soon speed up its own creation, that is not hype. It is a disclosure.
Sources: TechCrunch (first reporting on the filing), and Anthropic's amended Form S-1 (S-1/A) filed with the US Securities and Exchange Commission on 2 November 2026.
