Anthropic's IPO filing says it may soon cross its own threshold for AI-accelerated AI research

Most coverage of Anthropic's amended IPO prospectus has focused on its finances, which we covered yesterday. But the S-1/A filed with the SEC on November 2 also adds a new passage to its risk factors that TechCrunch spotted and read as a warning that the company is close to recursive self-improvement (RSI), the point at which AI systems speed up the development of AI itself.

What the filing says

The new language reads:

"For 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.' We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data."

The same section says that "much of our evidence regarding the pace of such AI-driven acceleration is derived from lagging indicators", that Anthropic's ability to forecast or measure its rate of progress "is limited", and that "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." Anthropic adds that it is "evaluating the implications of meeting the AI research and development capability threshold" and will release further reports as its policy requires.

According to Gizmodo, the October 26 version of the prospectus already discussed the rate of "AI-driven acceleration"; the "imminently" passage is what Monday's amendment added.

Three points are worth keeping straight:

• No model has crossed the line. The filing says so explicitly, and the forward-looking estimate concerns future models.
• This is a risk factor in a securities filing. Companies disclose risks to protect themselves legally. TechCrunch notes that an S-1 "functions as both disclosure and marketing".
• Anthropic does connect it to RSI. Elsewhere in the filing it says AI-assisted development is covered by its Responsible Scaling Policy, "which sets out our plans to monitor and mitigate the risks as we approach what has been termed 'recursive self-improvement.'"

What the threshold actually is

The phrase comes from Anthropic's Responsible Scaling Policy. In version 3.4, effective July 8, 2026, the automated AI R&D threshold is met if either:

1. Anthropic's models could "fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs (i.e., within a factor of 5)"; or
2. there is "dramatic acceleration" of the pace of AI progress for reasons that likely relate to the automation of AI R&D.

The second case is defined as observing or expecting "double the rate of progress" in aggregate AI capabilities, compared with both the expected rate and the fastest sustained rate seen without significant AI contributions, measured over at least three model generations, and plausibly attributable to automated research or engineering rather than, for example, more staff or compute. The policy gives a worked example: if baseline progress were a 9x "effective scaleup" per year, doubling the rate would mean something like 81x.

The policy's own changelog describes this threshold as "intended to capture the onset of dramatic recursive self-improvement", and admits it "has proven difficult to operationalize." It is the second scenario, not the first, that the S-1 language refers to.

If the threshold is met, the policy lists the measures Anthropic says it will take, including "moonshot R&D for security" projects "to explore ambitious and possibly unconventional ways to achieve unprecedented levels of security", and reaching an "eyes on everything" state for internal AI development, with comprehensive logging, internal compartmentalization and code review to limit opportunities for sabotage by AI models.

Claude building Claude

The context for the warning is spread across the prospectus. Anthropic calls the dynamic "closing the loop": successive Claude models help build the next one, including "running experiments, generating and processing training data, automating evaluations and safety audits, and writing and reviewing code", according to the filing.

It gives figures from its own engineering organization, based on internal benchmarks:

• Claude's success rate on "challenging, open-ended engineering problems that lack a clear specification" reached about 90% in September 2026, up about 50 percentage points in nine months;
• the average daily volume of code contributed per engineer in the three months to September 30, 2026 was over 45 times the 2024 average;
• over 90% of contributions to Anthropic's codebase in September 2026 were written by Claude.

These are company-reported numbers, and the filing does not publish the benchmarks behind them.

Anthropic also sets a limit on how far it intends to go. "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 writes, adding that it is "considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic."

Why it matters for developers and builders

For teams building on Claude, the most direct consequence is pace. The same filing says model versions released in the six months to September 30 made up more than 80% of third-quarter revenue. If AI-assisted research does double the rate of progress, model generations would arrive faster, and integrations pinned to one model would need re-testing more often.

The second consequence is policy. Meeting a Responsible Scaling Policy threshold triggers stricter internal safeguards and new risk reports, and Anthropic says it is weighing changes to the policy itself. Those decisions can affect release timing, access to the most capable models and usage terms. The company has not said when it will publish its assessment.

The third is industry-wide. The statement sits in a document filed with securities regulators, where misleading statements carry legal liability, rather than in a blog post or a keynote. Coming the same day OpenAI described rebuilding its Git backend to cope with the volume of code its Codex agents produce, it is a reminder that the labs now report their own AI as a main driver of how fast they ship.

What to watch: the risk report Anthropic says it will release, any revision to the Responsible Scaling Policy, and whether competitors preparing their own listings make similar disclosures.