When Anthropic unveiled its Long-Term Benefit Trust in September 2023, it framed the structure as a "corporate governance experiment." Three years later, the experiment is about to meet its first real-world exam: a public listing that could value the company near $2 trillion, according to estimates reported by the Financial Times. The question that matters isn't whether the governance architecture exists on paper. It's whether it can actually bind — the company, its founders, and its future shareholders — when the incentives to wriggle free are at their strongest.

What the structure actually is#

Anthropic is incorporated as a Delaware Public Benefit Corporation (PBC). That legal form gives directors the latitude to balance shareholder returns against the public-benefit purpose stated in the corporate certificate — in Anthropic's case, "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."

On top of that sits the Long-Term Benefit Trust (LTBT): an independent body organized as a "purpose trust" under Delaware common law. The Trust holds a special class of stock — Class T shares — that carries no economic value but grants trustees the authority to elect and remove an increasing number of Anthropic's seven board members, ultimately reaching a majority. Trustees are designed to be financially disinterested: no equity, no options, no bonuses tied to valuation. They receive flat compensation for their time. Trustees serve one-year terms and elect their own successors.

The Trust's authority phases in on time- and funding-based milestones, with a full board majority arriving within four years of the 2023 announcement. Protective provisions require the Trust to receive advance notice of actions that could materially alter the company.

Finally, there's a second charter of sorts: the Claude Constitution. In January 2026, Anthropic published a rewritten, roughly 23,000-word constitution for its models — an order of magnitude longer than the 2,700-word 2023 original — released under a Creative Commons CC0 public-domain license. Unusually, it is written primarily for Claude itself to read: Anthropic's bet is that models need to understand why they should behave well, not just what to do, so they can generalize to situations the authors never anticipated. Its four ordered priorities: be broadly safe, be broadly ethical, comply with Anthropic's guidelines, and be genuinely helpful.

Why it's genuinely unusual#

No other major frontier lab has built anything quite like the LTBT. OpenAI's nonprofit board structure is arguably the closest analogue — and its November 2023 board crisis, when directors briefly fired Sam Altman, demonstrated the dangers of mission-focused governance without institutional guardrails. Anthropic's design deliberately tried to avoid that outcome: instead of a nonprofit board with sweeping powers, it created an independent trust with board-appointment authority that phases in gradually, paired with failsafe amendment rules.

The mechanism has shown signs of life. Reports this month noted that the Trust receives advance notice of major management decisions, including new model launches, and recommended a limited release for the "Mythos" model rather than a full launch — the closest thing to a documented exercise of its oversight role.

And the Trust's independence is real in a way that matters for financial incentives: the trustees hold no equity, so a decision to delay a product for safety reasons doesn't cost them personally. In standard corporate governance, directors are often paid in stock, aligning their incentives with shareholders. The LTBT severs that link by design.

Where the binding starts to fray#

For all its ambition, the architecture has at least four soft spots — several of which are becoming visible right now.

1. The Trust is understaffed at the worst possible moment. When announced in September 2023, the Trust had five trustees, including chair Neil Buddy Shah (Clinton Health Access Initiative), RAND's Jason Matheny, and ARC's Paul Christiano. As of August 2026, reporting places it at just three members: Shah, Richard Fontaine (Center for a New American Security), and Ben Bernanke, who joined in July 2026. The fourth seat opened on August 4, when Mariano-Florentino Cuéllar stepped down to become Anthropic's first Chief Global Affairs Officer. No successor has been announced. A governance body operating below capacity while it finalizes the structure it will carry into public markets is not an inspiring look.

2. Self-selection is a feature and a bug. Trustees elect their own successors by a vote of the trustees, in consultation with company leadership. Anthropic argues this insulates the Trust from shareholder pressure. Critics note it equally insulates the Trust from any other accountability: nobody outside the room can install or remove a trustee. The mechanism depends entirely on the quality and independence of the people inside it.

3. The structure supervises positions, not relationships. As one careful analysis put it this year, every lever the Trust holds applies to the board and, through the board, the CEO. A trust can elect directors; it cannot regulate who the CEO listens to informally. If influence flows through relationships rather than formal authority, the most sophisticated governance document in the world has nothing to attach to. The Trust's mandate is also explicitly limited: it is designed for "extreme events," not day-to-day commercial strategy. That's reassuring for investors — and a reminder that most of what a company does happens outside the Trust's jurisdiction.

4. The IPO is adding a parallel power center. In August 2026, The Information reported — with confirmation from Bloomberg and Reuters — that Anthropic plans to issue super-voting shares to Dario Amodei and its six other co-founders ahead of a listing targeted as early as late September. CEO Dario Amodei reportedly owns only about 2% of the company after years of dilutive fundraising, so a standard one-share-one-vote structure would leave founders vulnerable to outside shareholder pressure post-IPO. The proposed fix would let founders' enhanced votes override ordinary public shareholders on major decisions.

How that founder-share class interacts with the Trust's Class T authority has not been publicly disclosed. The Trust's Class T power is not subject to any sunset provision in any published description, so even if founder super-votes eventually expired, public shareholders might still face a board they cannot elect. Institutional investors pushed back against similar dual-class structures at SpaceX this year, with three pension funds managing $1 trillion in assets calling it "the most management-favourable governance structure ever." Whether Anthropic's mission framing buys it a different reception remains to be seen.

What public shareholders actually get#

The uncomfortable summary for anyone buying into the listing: Class A shares, one vote per share, economic exposure to the upside — and, per reporting, effectively no ability to influence the board or override strategic decisions. Anthropic is organized to subordinate shareholder returns to its mission, and the Trust is the mechanism that makes that commitment credible. But "accountability to a different standard" only works if that standard is enforced by somebody with leverage.

Consider the incentives. The trustees are paid to be the brake — they have nothing to lose if the stock falls. That makes them immune to short-term commercial pressure, which is the point. It also makes them immune to the feedback mechanism that punishes bad judgment in ordinary companies. The architecture redirects accountability from shareholders to the Trust; the Trust's accountability is to its own purpose and its own successor selection. It is accountability as a closed loop.

The verdict: architecture with untested load-bearing walls#

Does the public-benefit structure still bind? The honest answer is that it binds in the way it was designed to bind — and that design has never been tested under load.

The honest case for the architecture: it's real legal machinery, not a mission statement. The Class T shares exist, the board appointments have been made (four of seven directors were reportedly nominated by the Trust, including Reed Hastings and Vas Narasimhan), the advance-notice provisions have been exercised on model releases, and the amendment rules create genuinely high barriers to dismantling. The Claude Constitution, published openly under a public-domain license, adds a layer of model-level governance no competitor has matched in transparency.

The honest case against: the Trust's actual enforcement power is untested and, per some assessments, uncertain. It sits at three of five seats while preparing for the IPO. The founders' super-voting plan introduces a parallel concentration of control whose relationship to the Trust is undisclosed. And the deepest critique — that the structure governs formal positions while real influence moves through informal channels — isn't a bug anyone can patch with another amendment.

The charter binds as far as the paper says it does. Whether it binds when it matters is the one question the next six months will answer.

Disclosure: This is an independent editorial analysis. No sources were interviewed for this piece; all facts are drawn from the public documents listed above. Reporting on the IPO structure and founder share plans is based on press reporting from The Information, Bloomberg, Reuters, and the Financial Times, as cited in the linked articles, and Anthropic has not publicly confirmed those details.