Anthropic's IPO filing warns recursive self-improvement could be "imminent"

What's going on

On Monday (November 2, 2026), Anthropic filed a revised version of its IPO prospectus (S-1/A) with the US Securities and Exchange Commission. The updated financials got the first attention. But the "Risk Factors" section contains a new passage that TechCrunch calls "a small bombshell in the fine print": Anthropic could soon reach its own threshold for recursive self-improvement (RSI). That means an AI that improves itself, with the improved version then improving itself further.

What Anthropic actually writes

Translated, the key sentence reads: internal estimates and expectations for future models, not current ones, suggest the company could "imminently" meet the definition of "dramatically accelerated progress attributable to AI" in its Responsible Scaling Policy (RSP, Anthropic's self-imposed safety rulebook). No current model has been classified as crossing this threshold, and the company says it continues to review the underlying data.

Other statements follow:

• Growing use of AI to speed up its own research could cause model capabilities to grow faster than the ability to develop, test and implement suitable safety, alignment and security measures.
• In the past, Anthropic only noticed a significant acceleration of its own development after it had happened, and it could do so again. According to QbitAI, that's because much of the evidence for research acceleration consists of lagging indicators.
• According to TechCrunch, Anthropic has already used AI help to build models whose jumps in capability it initially failed to recognize.

The prospectus also describes Claude as "one of the most powerful drivers" of Anthropic's own progress. Each new Claude version speeds up the next, for example through experiments, generating and preparing training data, automated evaluations and safety audits, and writing and reviewing code. The company calls this recursive dynamic "closing the loop." One expression of it, it says, is the growing pace of "compute multipliers," meaning improvements that get more out of the same computing power.

What RSI means and what the RSP threshold looks like

RSI is a fuzzy term that, according to Gizmodo, grew out of Eliezer Yudkowsky's older "Seed AI" concept. Pessimists see it as the point of no return where AI escapes human control. Optimists see the promise of enormous productivity and scientific discovery.

The RSP sets a more concrete yardstick, according to IBTimes: a doubling of the baseline rate of progress that can plausibly be attributed to AI. QbitAI writes that Anthropic treats AI development from 2018 to 2024 as the "1x baseline" and now considers it possible that AI could achieve a year's worth of progress equal to two baseline years. Gizmodo reads the RSP as saying that very high security standards would apply once the threshold is met. The company would explore "ambitious and possibly unconventional ways" to reach unprecedented security, establish a kind of total overview of internal activity ("eyes on everything"), and get so good at red-teaming (deliberate attack testing) that bug bounty programs would no longer be needed.

Humans stay in the loop, supposedly

Anthropic stresses that it does not expect to be able to safely let Claude autonomously design and build successive generations of frontier models, without meaningful human oversight or understanding, should that become technically possible. It is therefore considering changes to the RSP to keep humans in control of frontier development.

The financials in the background

According to QbitAI, the filing also includes quarterly figures: Q3 revenue of around $17 billion (above the previously stated estimate of $16.8 billion) and $33.24 billion cumulatively for nine months. Net loss was $23.29 billion in Q3 and $104 billion for nine months. QbitAI says most of that came from valuation changes on convertible notes and derivatives, not burned cash. The outlet puts compute commitments at $895.4 billion, with suppliers including Amazon, Google, SpaceX and so-called neoclouds.

QbitAI also points to a flattening: Q1 48 billion (+132%), Q2 116 billion (+144%), Q3 about 170 billion (+47%). Those figures are given there as dollar amounts of the magnitude stated in the article and are partly derived from run-rate estimates, not official monthly figures. Note: this comes from a single source and includes the outlet's own calculations. Anthropic itself stresses that run rate is not the same as actual revenue.

Extra claims, to be treated with caution

Financial Express reports that Claude agents write over 80% of the merged code in Anthropic's repositories, and that the speedup in training-loop optimization rose from 3× in mid-2025 to 52×. The other sources don't back up these numbers, so treat them as unconfirmed.

Our take

It's notable that a company ahead of an IPO is writing, in a legally liable document, that a central risk scenario of the AI safety debate may be close. At the same time, TechCrunch and IBTimes point out that an S-1 is both disclosure and marketing: hinting that you're close to self-building AI also signals technical strength. According to TechCrunch, it's undisputed that Claude has improved sharply this year in coding, math and cybersecurity. How reliable the internal estimates are remains open, since Anthropic itself admits it recognizes acceleration late. The announced RSP changes for human control would be a concrete thing to watch.