Formal structure. Anthropic is a Delaware Public Benefit Corporation whose specific public benefit is "to responsibly develop and maintain advanced AI for the long term benefit of humanity." Its signature mechanism is the Long-Term Benefit Trust (LTBT), an independent body of financially disinterested trustees holding a special Class T stock that elects an increasing share of the board over time. Initial trustees included Jason Matheny (RAND CEO), Neil Buddy Shah (chair, Clinton Health Access Initiative), Paul Christiano (Alignment Research Center), Kanika Bahl, and Zach Robinson. As of 2026, the Trust skews toward national security/policy: Richard Fontaine (CNAS CEO, ex–Defense Policy Board) joined in June 2025; Mariano-Florentino (Tino) Cuéllar (Carnegie Endowment president) in January 2026; founding members Bahl and Robinson departed.
On April 14, 2026, the LTBT appointed Novartis CEO Vas Narasimhan to the board, pushing Trust-appointed directors to a majority — a threshold written into Anthropic's founding documents but not exercised until then. The board also includes Dario Amodei, Daniela Amodei, Yasmin Razavi, Jay Kreps, Reed Hastings, and Chris Liddell.
Robustness/criticism. The structure has real but bounded teeth. "Failsafe" provisions let stockholders change the Trust and its powers without trustee consent if sufficiently large supermajorities agree (thresholds phase up over time). Critically, the Trust Agreement is enforceable by the company and by long-term stockholders — but not by the trustees themselves, a design choice critics (AI Lab Watch's review of Anthropic's Certificate of Incorporation) say undermines trustee independence. The Trust was also slow to use its powers (the first election, Jay Kreps, was announced May 2024, later than planned).
Track record. Anthropic published the first Responsible Scaling Policy (RSP) in 2023, built on AI Safety Levels (ASL); the current version is RSP 2.1 (effective March 31, 2025), with Capability/Safeguard Reports and a Responsible Scaling Officer. It has consistently held safety lines under commercial and now state pressure (see Pentagon and Fable/Mythos sections). The chief critiques are that its safety framing coexists with aggressive commercialization (Claude for Healthcare, a ~$1 trillion valuation, confidential IPO filing in June 2026, a SpaceX compute partnership) and deep national-security entanglement.
Formal structure. Founded 2015 as a nonprofit; created a capped-profit subsidiary in 2019; completed recapitalization on October 28, 2025 into OpenAI Group PBC controlled by the renamed OpenAI Foundation. Per OpenAI's own structure page, "as of the closing of the recapitalization, the OpenAI Foundation holds a 26% equity stake in OpenAI Group, worth approximately $130B"; Microsoft holds "roughly 27%" (~$135B), "and the remaining 47% is held by current and former employees and investors." The deal extended Microsoft IP rights to 2032, removed Azure exclusivity (in exchange for a $250B Azure commitment), and shifted AGI declaration from OpenAI's board to an independent expert panel. The California and Delaware AGs extracted ~20 concessions (the PBC must prioritize mission over shareholders on safety/security; the Foundation can halt model releases).
Robustness/criticism. Critics (Public Citizen, Eyes on OpenAI coalition, ex-employees via Lessig amicus brief, the Midas Project) argue the nonprofit's control is largely nominal — boards are nearly identical, profit caps were removed, and the structure was driven by a SoftBank funding condition requiring conversion. Musk's suit to reverse the conversion went to a jury, which on May 18, 2026 found for OpenAI/Altman/Brockman/Microsoft.
Track record — the worst under pressure. The November 2023 board crisis: the nonprofit board fired Altman ("not consistently candid"); Helen Toner later cited his failure to disclose ChatGPT's launch and his ownership of the OpenAI Startup Fund, plus executive reports of "psychological abuse." Within two days the revolt was overwhelming: by midday Monday, Nov. 20, 2023, more than 700 of OpenAI's ~770 employees (over 90%) had signed the open letter, per Axios ("By midday Monday letter organizers said the number of employees who'd signed the letter had grown to more than 700"); the letter read, "We will take this step imminently, unless all current board members resign." Microsoft backed Altman, and he returned while the dissenting board members (Toner, McCauley, Sutskever) exited. The 2024 Superalignment dissolution: the team promised 20% of compute was disbanded after Sutskever and Jan Leike left ("safety culture and processes have taken a backseat to shiny products"); Fortune reported the 20% compute was never delivered. In February 2026, OpenAI disbanded its Mission Alignment team after 16 months. This is a repeated pattern of safety infrastructure being subordinated to commercial velocity.
Formal structure. Formed April 2023 by merging DeepMind and Google Brain; a wholly owned Alphabet subsidiary, with Demis Hassabis reporting to Sundar Pichai. The original DeepMind independent-governance ambition (the "Project Mario" 3-3-3 board, the "global interest company" structure, the DeepMind Ethics & Society unit, and the AI ethics board promised at acquisition) was never realized; the merger ended DeepMind's push for autonomy. Internal safety governance now runs through the Responsibility and Safety Council (RSC) and the AGI Safety Council (led by Shane Legg), with a Frontier Safety Framework (v2.0) defining Critical Capability Levels.
Robustness/criticism. There is no mission-lock independent of Alphabet's commercial control. In February 2025, Google removed its 2018 pledge not to build AI for weapons/surveillance, reframing around "democracies should lead." DeepMind also fired 28 employees over 2024 Project Nimbus protests; in 2025–26, ~1,000 London staff became the focus of a unionization drive with the CWU/UTAW over military contracts (Project Nimbus, a $1.2B Israel cloud deal; a US DoD classified deal). Per the CWU (reported by Fortune, May 5, 2026), 98% of its DeepMind members backed the bid for recognition of the Communication Workers Union and Unite; Google DeepMind responded, "At this stage in the process, there has been no vote to unionize." At least five resignations and threatened research strikes have accompanied the dispute.
Track record. Strong scientific/safety research output (AlphaFold Nobel, interpretability) but governance subordinated to Alphabet; commercial pressure (Gemini integration) and military contracting have repeatedly overridden the original ethics commitments.
Canadian (Toronto), founded 2019 by Aidan Gomez, Ivan Zhang, Nick Frosst; ~$1.54B raised, ~$7B valuation (Sept 2025 $100M close); investors include Inovia, Radical Ventures, AMD, Nvidia, Salesforce, PSP, Oracle, plus Canadian pension/development capital. A conventional VC-backed C-corp with a five-member board; no AGI safety commitments, no RSP, no mission-lock. Strategy is "security-first," sovereign/enterprise AI (North platform), with growing defense work (Thales naval, Saab GlobalEye, Hanwha Ocean) and a planned acquisition of Germany's Aleph Alpha (April 2026). Governance question is essentially moot: Cohere does not claim to be building AGI or to need AGI-specific guardrails.
French, founded 2023 (Arthur Mensch, Guillaume Lample, Timothée Lacroix). Series C September 2025 led by ASML (€1.3B), valuing it ~€11.7B; ASML is the largest shareholder with a strategic-committee seat. Other backers: Nvidia, DST, a16z, Bpifrance, General Catalyst, Index, Lightspeed. Governance philosophy is open-weight + European sovereignty, not AGI-safety mission-lock. Mistral, like 100+ companies, urged the EU to delay AI Act implementation; Mensch positions on user control and not being "dependent on external providers that can turn off that button." Macron has championed it ("download Le Chat"). Strengths: open-weight transparency, jurisdictional/sovereignty differentiation. Weaknesses for AGI governance: open-weight release is irreversible and a misuse liability; no catastrophic-risk framework comparable to RSP/FSF.
Microsoft AI is led by Mustafa Suleyman (DeepMind co-founder), installed via the March 2024 Inflection acquihire — a ~$650M "licensing" deal that hired Suleyman, Karén Simonyan, and most staff without a formal acquisition, drawing an FTC probe into whether it was structured to evade antitrust review (companies must report acquisitions valued over $119M). Microsoft's governance posture is that of a conventional public company with no AGI-specific mission-lock; its main AGI exposure is contractual (the OpenAI partnership, now able to pursue AGI independently post-2025 deal). Microsoft also has substantial defense work. The Inflection structure itself is a governance red flag: a template for "absorb the talent, avoid the review."
This is the central dynamic of 2025–26. Government action is overriding and reshaping internal lab governance through procurement, export controls, and national-security directives.
Pentagon/defense entanglement. On July 14, 2025 the DoD's Chief Digital and AI Office (CDAO) awarded contracts valued at up to $200M each to Anthropic, Google, OpenAI and xAI; CDAO chief Dr. Doug Matty said, "The adoption of AI is transforming the Department's ability to support our warfighters and maintain strategic advantage over our adversaries" (Nextgov/FCW). Anthropic was first to put a frontier model on classified networks (via the Palantir/AWS partnership) and Claude was reportedly used in the Maduro capture and to help identify ~1,000 targets in Iran strikes. The Anthropic–Pentagon rift is the defining case: Anthropic refused the "all lawful use" standard, insisting on carve-outs against mass domestic surveillance and fully autonomous weapons. The administration responded by designating Anthropic a "supply chain risk" (Feb 27, 2026 executive order), historically reserved for foreign adversaries; Trump ordered federal agencies to phase out Claude; Treasury, State, and HHS directed staff off Claude; defense contractors (Lockheed) began removing it. Anthropic sued, won a preliminary injunction in California, but an appeals court declined relief on the supply-chain-risk designation. OpenAI capitalized, signing its own classified Department of War deal (Feb 27, 2026), which Altman admitted was "rushed"/"sloppy"; OpenAI emphasized retaining its safety stack, cloud-only deployment, and cleared personnel. xAI simply accepted "all lawful use." This shows that the government can punish the most safety-restrictive lab and reward the most permissive — a structural incentive against internal restraint.
Google's military history. Project Maven (2017), the resulting employee revolt and the 2018 AI principles, then the February 2025 reversal of the weapons pledge, plus Project Nimbus and the 2024 firings, show the same erosion.
Export controls and compute governance. The Biden AI Diffusion Framework (Jan 15, 2025) — which for the first time controlled model weights alongside chips and created country tiers — was rescinded by the Trump administration on May 13, 2025 as "divorced from commercial reality." BIS issued interim guidance (presumptive violations for dealings in Chinese advanced ICs like Huawei Ascend). The administration then pivoted to bilateral, transactional diffusion: Trump's December 8, 2025 announcement permitting some chip sales to approved China customers, and the January 13, 2026 BIS case-by-case licensing rule. Chips are now bargaining chips. This interacts with lab governance: a permissive diffusion regime favors closed-but-widely-sold models and undercuts the open-vs-closed safety calculus.
This is the single sharpest illustration of state power overriding lab governance.
Background — Mythos and Project Glasswing. Claude Mythos is Anthropic's most capable, most restricted model, strikingly strong at finding software vulnerabilities. Project Glasswing (launched ~April 2026) gave ~50 vetted partners (AWS, Microsoft, Nvidia, Google, CrowdStrike, Palo Alto Networks, JPMorganChase, the Linux Foundation, US government) access to Claude Mythos Preview to scan critical codebases; partners found 10,000+ high/critical vulnerabilities, and Anthropic's own scans of 1,000+ open-source projects flagged 6,202 high/critical (90%+ of a reviewed sample validated). Anthropic committed up to $100M in usage credits and $4M to open-source security, and explicitly declined to release Mythos-class models publicly for lack of sufficient safeguards. On June 2, 2026, Glasswing expanded to ~150 organizations across 15+ countries — the same week Anthropic filed confidentially for IPO and the Trump administration signed its AI security EO. Fable 5 is the public, safeguarded model built on Mythos technology; Mythos 5 is the unsafeguarded full version for trusted partners.
The directive. On June 12, 2026 at 5:21 PM ET, Anthropic received a Commerce Department export-control directive (a letter from Commerce Secretary Howard Lutnick, drafted with BIS) to suspend all access to Fable 5 and Mythos 5 by any foreign national, including Anthropic's own foreign-national employees. Because nationality cannot be filtered in real time, Anthropic disabled both models for all customers worldwide — reportedly the first time a leading lab took a publicly deployed model offline due to direct federal intervention. The asserted trigger was a reported "jailbreak"; Anthropic reviewed the underlying report and said the capability was narrow, non-universal, "widely available from other models (including OpenAI's GPT-5.5)," and used daily by defenders. Per Bloomberg, the Lutnick letter required a license before export to any foreign national, threatened criminal/civil penalties, and cited concerns about use by military intelligence in China and Russia.
What it reveals about Anthropic's governance. Anthropic complied with the legal directive while publicly and sharply disagreeing: "we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers." It invoked its stated principle that government should be able to block unsafe deployments only "as part of a statutory process that is transparent, fair, clear, and grounded in technical facts," and said this action did not meet that bar. This is consistent with Anthropic's broader pattern — it accepts national-security framing and works closely with government, but contests overreach through institutional channels (statements, litigation) rather than quiet capitulation. It also reveals the limits of even the strongest internal governance: the LTBT, RSP, and board majority are irrelevant when a Commerce directive can pull models in hours.
What it reveals about government as a governance force. A single directive took a generally available product offline for its entire global user base within hours. As of mid-June 2026, access had not been restored; Anthropic reported working with the White House to end the restrictions. A "FreeFable" open letter organized by former Facebook/Yahoo security chief Alex Stamos (addressed to Lutnick and National Cyber Director Sean Cairncross) urged rescission, drawing 120+ signatories from across the security industry. Security analysts (Snyk, Cloudflare) drew the operational lesson: availability you don't control is a risk, favoring open-weight/self-hosted models — a direct feedback into the open-vs-closed governance debate.
On June 2, 2026, Trump signed an executive order, "Promoting Advanced Artificial Intelligence Innovation and Security," establishing a voluntary framework under which frontier developers may give the government access to a "covered frontier model" for up to 30 days before release (cut from 90 in earlier drafts). The order explicitly bars any "mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models." Treasury, the NSA (which designates "covered frontier models"), and CISA run the review, with NIST/Commerce input; the framework is due ~August 1, 2026. Industry leaders broadly welcomed it (Altman called the balance "appropriate"; Anthropic said it "looked forward to working with the administration"), while Senator Josh Hawley wanted mandatory reporting. Around June 4–5, 2026, Anthropic separately published a report proposing that major AI companies in the US and China coordinate a verifiable pause/slowdown: "We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology." Dario Amodei's June 10 "Advanced AI Framework" called for mandatory third-party testing above a compute threshold (cybersecurity, bioweapons, loss of control, automated R&D) and government power to block unacceptable-risk deployments, while doubting China would be a responsible pause partner.
Anthropic comes out best overall — it has the strongest mission-lock, the best safety track record under both commercial and state pressure, and the highest structure-values coherence. But three caveats are essential. First, Anthropic's advantage is relative, not absolute: its governance can be overridden in hours by a Commerce directive, its Trust has enforceability gaps, and its commercial trajectory (trillion-dollar valuation, IPO, defense revenue) creates the same erosion pressures that have degraded peers. Second, the most important governance variable is no longer corporate structure but government action — procurement, export controls, and national-security directives now dominate, and they currently punish safety-restrictiveness (Anthropic blacklisted) and reward permissiveness (xAI in classified systems). Third, on the question that matters most — handling a genuine AGI transition — all of these structures are inadequate to the stakes, designed for a world where the binding constraint is the company's own resolve, not a state that can commandeer or sideline the company.