AI writes 90% of the code: Anthropic warns of speed without control
In its amended IPO prospectus of 2 November, Anthropic says it may soon reach its own threshold for development dramatically accelerated by AI.
An AI developer preparing to list has written into a mandatory filing that its models may soon sharply accelerate its own research, and that it might recognise such leaps only after the fact. A safety risk thus becomes a matter of liability and valuation.
What happened
The disclosures appear in an amended registration statement, Form S-1/A, filed with America's Securities and Exchange Commission on 2 November. Internal estimates suggest the company "may imminently meet" the definition of "dramatically accelerated progress attributable to AI" in its own safety framework, the Responsible Scaling Policy. According to the prospectus, that means AI systems replacing the entire research workforce at competitive cost, or doubling the overall pace of AI progress. Anthropic has not designated any current model as exceeding the threshold. In September, the filing says, Claude authored more than 90% of contributions to the company's codebase, and daily code output per engineer in the third quarter was more than 45 times its 2024 level. That measures volume, not productivity. Anthropic concedes it has already used AI to build models whose capability leaps it did not recognise at the time. It does not expect to be able to let Claude safely build successive frontier models without human supervision, and is considering stricter rules. Assessment: a prospectus is both disclosure and marketing, as TechCrunch also notes.
What follows
If a laboratory notices capability jumps only after training, oversight shifts from before release to after it. For firms using AI models in customer service, software or financial processes, vendor testing alone counts for less, and their own checks before each model switch matter more. For investors a safety question becomes a valuation factor, because the risk language now sits in a document that carries legal liability. Assessment: whether the threshold will in fact be reached soon cannot be verified from outside. What is established is that a large developer formally describes the possibility.
What changes in practice
- Companies
- New model versions may prove far more capable than announced. In-house testing before switching matters more than vendor claims.
- Investors
- The risk of uncontrolled AI acceleration now sits in the prospectus and becomes part of the valuation.
- Software developers
- At Anthropic, Claude writes more than 90% of code contributions, the filing says; humans set goals and review.
What to watch
- Whether Anthropic designates a model as above the threshold
- Changes to the Responsible Scaling Policy
- Further prospectus amendments or SEC queries before the listing
Sources
- SEC: Anthropic PBC, Form S-1/A (2. November 2026)
- TechCrunch: Anthropic warns that it's very close to RSI capability
- Wccftech: Anthropic warns self-improving AI milestone may be imminent