The Largest IPO in History, Anthropic’s S-1 Reads Like an “Anti-Sell”

Yesterday, Anthropic filed a 318-page S-1 with the SEC, planning to list under the ticker ANT. Jerry Liu and his team used LlamaParse to read every page and built a two-column reader. Every number is annotated with a page number, and every quote can be traced back to the original text: https://www.documentstories.com/anthropic-s1

The page is called “The Anti-Sell.” The title comes from advice heard repeatedly: “Never let your company’s value exceed its actual value.” A prospectus is usually about selling a story; this one feels more like saying “what we have already achieved.”

Start with the numbers. Revenue went from $41 million in 2023 to $386 million in 2024, $4.585 billion in 2025, and $16.286 billion in the first half of 2026. Every quarter’s number keeps setting new records. Q1 2025 was still $376 million, while Q2 2026 had already reached $11.555 billion.

Then there are the losses. Net loss in the first half of 2026 was $80.7 billion. But this number needs to be broken down: $72.4 billion came from fair value changes in convertible notes, derivative liabilities, and SAFEs—when the valuation rises, a loss is recorded on the books. Operating loss was only $693 million. The company held $131.65 billion in cash and short-term investments.

In its S-1, Anthropic explicitly lists some “things we don’t do”: no advertising; no development of image and video generation models; possibly slowing down progress in exchange for safety; giving about $100 million in Claude credits to cyber defenders. The co-founders have each committed to donating 80% of their personal equity.

The governance structure is also unusual. Seven founders permanently hold 50.1% of voting power through Class F. Of the seven board seats, four are selected by a trust called the Long-Term Benefit Trust, whose trustees include Ben Bernanke. The prospectus itself admits: “This structure has no direct precedent.”

There is one more detail: more than 90% of code contributions are written by Claude, and engineers’ average daily code volume is 45 times the 2024 average. The company calls this a “closed loop”—AI helping people write AI. The risk factors also acknowledge that this could cause the company to lose visibility and control over certain processes.

The reader itself is worth mentioning. The entire PDF has 421 pages; LlamaParse read it in 288 seconds and output 13,847 lines of markdown. Every number carries its source page number. In the comments, someone asked how to ensure the model does not make up numbers, and another user suggested: don’t let the model output numbers directly; have it point to table cells and retrieve values by coordinates. This tool’s approach is probably similar.

For ordinary readers, this page is much friendlier than the PDF. For developers, it demonstrates a reliable document parsing pipeline.

Anthropic’s S-1 contains a lot of information, and this reader is only an entry point. The original is on the SEC website, and the bottom of the page also notes: not affiliated with Anthropic, not investment advice.