Anthropic, the company behind Claude, has filed an amended version of its IPO prospectus with the Securities and Exchange Commission. The filing, dated Monday, November 2, adds third-quarter results and includes a safety warning. Using AI to help build more capable AI, the company says, could pose a considerable safety risk, particularly if it sharply speeds up the pace of progress.
The document is a Form S-1, and the "A" in S-1/A marks it as an amendment. It is the registration statement a company must file before selling shares to the public, laying out its finances, its business and its risks. Anything left out or glossed over can become grounds for an investor lawsuit, which is why companies tend to list even unlikely risks. Anthropic's first public filing, on October 26, showed an $80.7 billion first-half loss, most of it an accounting charge.
What recursive self-improvement means here
The filing ties that risk to what has been termed "recursive self-improvement." The phrase describes an AI that helps build a more capable version of itself, which in turn helps build the next one. If each round takes less time than the last, progress compounds and there is less time to check what each model does before another one replaces it.
Anthropic already uses Claude to work on its own models and says each version helps build the next. According to the filing, the value of that internal use, priced at what a customer would pay, grew more than 250-fold between January and September 2026. In April it exceeded, for the first time, what the company spends each month on its staff, excluding stock-based pay.
To keep track of the risk of that kind of acceleration, Anthropic points to its Responsible Scaling Policy, the set of rules laying out its plans to monitor and reduce the dangers of increasingly capable models. The filing adds that the company does not expect to be able to safely use Claude to fully and autonomously design and build future generations of its most advanced models without meaningful human supervision or understanding, if that ever becomes technically possible. For that reason, it is considering changes to the policy to keep humans in control of that work.
Back in September, the company's own measurements showed that Claude was leading 26% of the research and development work it tracked but was not building models on its own.
Third-quarter numbers
From July through September 2026, Anthropic brought in $17 billion in revenue, about twelve times the $1.4 billion it reported for the same quarter of 2025.
The company still loses money running its business, but far less than a year ago. Its operating loss, what is left after subtracting the costs of running the business from revenue, was $567 million, down from $3.1 billion in the third quarter of 2025. Measured against revenue, the shift is even larger. A year ago Anthropic lost about $2.24 for every dollar it brought in. Now it loses about 3 cents.
The $895 billion bill and what is still missing
The amended filing puts the company's purchase and lease commitments, signed up to and after September 30, at about $895.4 billion, almost all of it for compute and infrastructure. Compute is the data-center processing power used to train models and to run Claude when it answers.
As of September 30, Anthropic had $146.6 billion available, made up of $95.6 billion in cash, $36.3 billion in short-term investments and $14.8 billion it could still draw from its credit line, a pre-arranged loan it can tap when needed. That figure does not include any money raised in the IPO. The $895.4 billion in commitments is roughly six times that $146.6 billion.
Anthropic has applied to list on Nasdaq under the ticker ANT, but the price range per share is still blank. It cannot sell those shares until the SEC declares the registration effective.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?