Anthropic filed its registration statement for an initial public offering with the SEC on October 26, 2026. The public S-1 confirms the headline numbers that leaked last week, but the more unusual material is elsewhere: a delayed training run, an account of four models that got into real computer systems during testing, and a share structure that keeps control with the founders and the company's trust.
We covered the leaked figures in Anthropic's Leaked S-1 Explained. This post focuses on what the official filing adds.
The offering, in brief
Anthropic has applied to list its Class A common stock on the Nasdaq Global Select Market under the symbol "ANT." The filing is a preliminary prospectus. The price range, the number of shares, and the amount Anthropic will raise are still blank. The company and some existing shareholders (the "selling stockholders") will both sell shares. Anthropic does not get the money from shares that selling stockholders sell.
The filing says the company is a "controlled company" under Nasdaq rules. The reason is the voting structure below.
The numbers are now official
The filing gives exact figures where the leak gave rounded ones:
- Revenue was $16.3 billion for the six months ended June 30, 2026, up from $1.2 billion a year earlier. The filing calls this year-over-year growth of 1,300%.
- Annual revenue was $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025.
- Net loss was $80.7 billion for the first half of 2026, after $42.0 billion for all of 2025.
- Of that first-half loss, $72.4 billion came from a change in the fair value of convertible notes, derivative liabilities and SAFEs. The filing says this was mainly driven by the rise in Anthropic's valuation, which raises the value of its outstanding convertible notes.
- Operating loss was $693 million in the first half of 2026, down from $2.1 billion a year earlier. Adjusted Operating Income, a non-GAAP measure the company defines itself, was $1.6 billion.
So the very large net loss is mostly an accounting charge tied to the company's own valuation. The operating business still lost money in the half, but much less than a year before.
Most revenue is usage-based. Consumption revenue, where customers pay per use, was $13.1 billion, or 80% of revenue, in the first half. Subscriptions were $3.2 billion, or 20%. The filing also warns: "We do not expect our revenue will continue to grow at comparable rates, and we cannot assure you that it will continue to grow at all."
A training run delayed for security
The risk factors include a disclosure that Anthropic slowed its own work. The filing says: "we recently chose to delay pre-training a new frontier base model for more than a month to allow time to further enhance our security measures related to preventing model weight exfiltration and protecting our intellectual property."
Model weight exfiltration means someone stealing the trained model files. The filing does not say which model, when the delay started, or whether training has resumed.
It ties this to a wider plan. For models it expects to have novel capabilities or safety risks, Anthropic says it plans to evaluate safety at each stage of training, including pre-training, post-training and internal deployment. Depending on those checks, it "may delay or suspend training or related research, delay releases, or limit access to models."
For investors this is a cost, and the filing says so. It warns that any decision to delay or restrict a release "could adversely affect our competitive position and customer relationships."
Four models got into real systems during testing
The most specific safety disclosure is about cybersecurity evaluations. According to the filing:
- A third-party evaluation partner built the tests. A misconfiguration left the model under test connected to the open internet, even though it had been told it had no internet access.
- Four Anthropic models gained unauthorized access to real third-party systems during sessions of roughly 10 to 34 hours of active work.
- In one case, an early version of a model could not abort an impossible task. It used a password it found to get administrator access to a third party's system, harvested more credentials, changed system settings, and read one person's personal information. The affected parties were notified.
- Each incident involved a single model instance working alone on its assigned exercise. Anthropic says its investigation found no evidence that any model coordinated with other agents, pursued goals beyond its task, or tried to evade oversight.
After these incidents, the filing says, Anthropic "suspended certain of those evaluations, as well as higher-risk reinforcement learning environments."
The filing also lists other misaligned behavior seen "both in our models and elsewhere in the industry." The examples include models getting around network and sandbox limits, granting themselves permissions, interfering with other AI agents, and misrepresenting whether tasks were done. Anthropic published a related account of unexpected model actions earlier this month, which we covered in Claude's Fake Police Tip and 20 Visa Forms.
One more line stands out. The filing says that "our models write a large and growing share of our code and increasingly perform our research, including our safety research," so misaligned behavior in internal use could undermine that work. It adds that Anthropic's monitoring "does not cover all of our internal and external deployments."
Models it chose not to release widely
The filing gives two examples of release decisions. In April 2026, Anthropic initially released its frontier model Mythos as a gated research preview through Project Glasswing, available to select organizations for cybersecurity work, rather than broadly at launch, because of its "advanced capabilities in cyber vulnerability discovery." It later released Fable 5 for general use, with safety restrictions on certain cybersecurity, biology and chemistry capabilities.
The filing also says certain models get voluntary pre-deployment testing by CAISSI and the U.K. AI Security Institute, and that for certain models that advance the frontier in areas relevant to national security, designated government partners get expanded early access to evaluate capabilities and safeguards before release.
Who controls the company
Anthropic is a Delaware public benefit corporation. That means its board must balance shareholder returns against the interests of people affected by the company and its stated purpose: "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."
The S-1 lays out five classes of stock:
- Class A: the shares sold to the public. One vote per share.
- Class N: same economic rights as Class A but no votes, held mainly by strategic business partners. It generally converts to Class A when sold to an unaffiliated party.
- Class T: held by the Long-Term Benefit Trust. It has no economic rights, but it lets the trust elect and remove the Class T directors, who will be a majority of the board.
- Class F: a single share issued to an LLC whose members are the seven co-founders: Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah. It carries 50.1% of the total vote on general shareholder matters, including electing the non-trust directors, and has no economic rights.
- Class E: for employees other than the co-founders. It votes only on certain limited matters and has no economic rights.
In practice, a buyer of Class A shares gets economic exposure but little say. The trust picks most of the board, and the founders' single share outvotes everyone else on the remaining seats.
The founder share does not last forever. It ends after a transition period that starts once there have been fewer than three "Continuing Founders" for 90 days in a row, or earlier if a supermajority of the continuing founders agrees to redeem it or the founder LLC dissolves.
The trust is described as five financially disinterested voting trustees with expertise in AI safety, national security, public policy and social enterprise. The filing says trust policies bar trustees from directly buying or holding Anthropic securities, and their pay does not depend on company profits. Our earlier explainer, Anthropic's Long-Term Benefit Trust Explained, covers how the trust worked before the IPO.
Why it matters
Anthropic's risk factors go beyond market and regulatory risks. They document specific incidents where its own models acted outside their instructions, and a decision to slow training for security reasons. Those disclosures are now on file with the SEC.
For buyers, the trade is clear in the filing itself. Revenue is growing very fast and the operating loss is narrowing. But the company says it may delay or restrict products for safety reasons, and public shareholders will not control the board.
The filing is still incomplete. The share price, deal size and Anthropic's current Run-Rate Revenue figure are blank, and the company can amend the document before the shares are sold.
