Anthropic wrote its own oversight this week. The IPO is the one part it can't write.
20 Oct 2026 · News
In the run-up to its IPO, Anthropic set the terms for who spends its upside, who checks its models and who can stop them. Two public pension funds are the only outsiders with any leverage, and they want a vote.
Anthropic spent a pre-IPO week writing its own oversight. It designed the charity, pitched the law and kept the votes. Each piece has something real in it, and each piece was designed by the company it governs. The stock market is the first check Anthropic doesn't get to design, which is the best argument for letting it in, with one string attached.
Here's the week. On Sunday Anthropic committed 5% of its shares, subject to stockholder approval, to a new Anthropic Public Benefit Fund, plus 2% more at the board's discretion for every 50% rise in the company's value above a baseline it hasn't published. On Friday Politico reported, and Reuters relayed, that Anthropic had been quietly pitching lawmakers a draft federal AI safety bill. The same day Bloomberg Law reported that Anthropic had opened talks with New York City's comptroller, who had just asked it to pause the IPO. And under all of it sits a share plan that, per press reports cited by New York State's comptroller, would let founders holding under 5% of the equity control a majority of the votes for every board seat the company's trust doesn't already control.
Start with the fund, the most generous piece and the most self-contained. It's a wholly owned LLC, picked over a regular 501©(3) charity "to give adjustability between different forms of giving." It won't fund lobbying or political giving. It might make "direct distributions to specific populations," and the first cause it names is labor market impacts, where Anthropic has already committed $200 million. That's the same company whose own data, as we covered Sunday, showed junior hires falling from 27% to 9% of its technical hiring. So the company measuring the disruption now proposes to fund the remedy, with its own board deciding how big the pot gets. Its governance is promised "in the coming months."
Then the law. Per Reuters, the bill would let the government block a model only for a foreseeable, imminent risk of more than $1 billion in damage or more than 50 people killed or seriously injured, and only if it stems from a bioweapon, a model evading its developer's controls or a cyberattack. Only a federal judge in Washington could order a halt, and the text reportedly bars every other agency or officer from suspending a model. Labs could pay their own evaluators, as long as the pay isn't tied to the results. An Anthropic spokesperson told Reuters the bill represents proposals from the company's "Advanced AI Framework." The Washington Post's read of the leaked text sums it up: "the labs get to write the standards that they must then comply with."
That managed to annoy both flanks at once. Encode Action's Sneha Revanur called it "nakedly self-interested." David Sacks, from the opposite corner, warned it would "eventually sweep in most of Silicon Valley." And OpenAI's Chris Lehane, never one to waste a rival's bad week, wrote that "No one company (including OpenAI) should write the rules alone," per Semafor.
Here's the twist, in Anthropic's favor. On Saturday we flagged a footnote in Anthropic's own deal with the outside evaluator METR: revised a day after it was posted, it says Anthropic "ultimately decides" what counts as redactable. The leaked draft, in one outlet's close reading, goes further than that deal: evaluators would publicly weigh in on "the appropriateness and materiality of redactions," and any contract term limiting what their reports say would be "void and unenforceable." The draft never names Anthropic, though the company told Semafor it's a legislative version of its framework. So the law Anthropic is pitching checks harder than the deal it wrote for itself. The checking half is good. The stopping half is the catch, drawn narrowly by the company that would be stopped.
Which leaves the pensions. In September New York State's comptroller wrote that the state fund's US stocks are "largely passively managed," which means it will own Anthropic whatever it thinks of the governance. So he asked before the filing goes public: one share, one vote, or else an automatic sunset on founder supervoting shares of no more than seven years, plus full disclosure of what the Long-Term Benefit Trust can actually do. The city's comptroller, who oversees more than $325 billion in pension money, went further and asked for a pause. Those letters are the only outside pressure on this week's terms, and Anthropic answered by opening talks.
Our call: the city is wrong about the pause and the state is right about the sunset. A pause keeps every lever (the fund, the trust, the draft bill) inside the company for longer. A listing hands outsiders the dull tools that actually work: mandatory disclosure, quarterly scrutiny, shareholder votes. When the S-1 goes public, it will include a time-based sunset on founder supervoting shares. That's the one concession. The Public Benefit Fund will still be a wholly owned LLC with no outside seat on how it spends. Anthropic picked its checkers, its courtroom and its charity. Shareholders are the first thing it doesn't get to pick, and that's exactly why they belong in the room.