A safety-first lab that is now a near-trillion-dollar commercial bet
Anthropic is the clearest live test of whether a lab founded to put safety research at the center of frontier AI can still win commercially. It is a San Francisco public benefit corporation whose stated purpose is the responsible development of advanced AI for the long-term benefit of humanity, and it sells that research as Claude: models and agents used in chat, coding, office work and enterprise deployments.
The commercial scale is no longer hypothetical. A prospectus reviewed by Reuters, as reported by CNBC, shows 2025 revenue of nearly $4.6 billion, about twelve times 2024, against a $42 billion net loss and more than $8 billion of operating losses excluding financing writedowns. Anthropic declined to comment on the documents. The same reporting places a May 2026 valuation at $965 billion and says a public listing is likely to slip until after the November 2026 U.S. midterm elections.
Former OpenAI researchers, a public benefit charter and a trust that shares the board
Anthropic was founded by researchers who left OpenAI after disagreements over governance and AI safety, including CEO Dario Amodei. It released its first large language model in March 2023. The current board, as listed on the company site, is Dario Amodei, Daniela Amodei, Yasmin Razavi, Reed Hastings, Chris Liddell and Vas Narasimhan.
Alongside ordinary stockholder elections, Anthropic's Long-Term Benefit Trust also elects directors. The company currently lists three trustees: Neil Buddy Shah, Richard Fontaine and former Federal Reserve chair Ben Bernanke. That structure is the governance expression of the same bet as the product line: keep a public-benefit mission enforceable as commercial pressure rises.
Constitutional training, encoded refusals and a Pentagon loss
Anthropic's technical identity is that safety is built into the model, not only bolted on afterward. The D.C. Circuit opinion describes three layers: training Claude against a constitution of principles, technical monitors stacked on the model, and contractual usage limits. Those encoded refusals are why Claude can decline a narrow set of high-risk cyber or biology requests on Sonnet 5.5, and why the Pentagon treated the same design as a supply-chain risk.
On September 25, 2026, a split D.C. Circuit panel rejected Anthropic's challenge to the Department of War's decision to exclude Claude from its supply chain under the Federal Acquisition Supply Chain Security Act. The majority held that the department had ample support for finding a national-security risk after Anthropic refused to drop contractual bans on lethal autonomous warfare and domestic surveillance. CNBC, summarizing the political backdrop, notes that a U.S. judge blocked a broader Pentagon blacklisting in August; the September appellate ruling still lets the Defense Department keep Claude out of its systems and contractor work.
From a chatbot family to Claude 5.5 and a deployment academy
The prior overview still described a Fable, Opus, Sonnet and Haiku stack from early September. That is now stale. Anthropic rolled out Opus 5.5 the week before September 28, which CNBC casts as a response to OpenAI's GPT-6 Astra, and on September 28 it shipped Sonnet 5.5 as the second Claude 5.5 model. Sonnet 5.5 keeps Sonnet 5's $2/$10 per-million-token list price, runs 30%+ faster, and is the everyday complement to Opus 5.5's heavier judgment work. It is on AWS, Google Cloud and Microsoft Azure. Haiku 5.5 remains unreleased; Anthropic says it will join in the coming weeks.
The product is no longer only a model card. On October 2 Anthropic launched Claude Frontier Academy with a $100 million commitment to train 10,000 Frontier Deployed Engineers by the end of 2027, starting with cohorts from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk in San Francisco, New York and London. First Frontier Deployed Engineer badges are expected in early 2027. The program sits on a Partner Network Anthropic says already spans 46,000 firms and more than 175,000 Claude certifications.
A usage business sitting on half a trillion dollars of planned compute
Anthropic makes money when customers use Claude: API and cloud-platform tokens, consumer and enterprise seats, and agent tools that Frontier Academy is meant to push deeper into large organizations. The prospectus, as reported by CNBC, warns that nearly a quarter of 2025 revenue came from two customers and that many large clients are not locked into long-term contracts.
The cost side is compute. The same documents say Anthropic spent $7.33 billion on compute and infrastructure in 2025, more than half of $12.65 billion in operating expenses, and plans $518 billion of cloud, computing and infrastructure obligations in coming years. That is the physical constraint on both the safety story and the IPO story: the lab cannot stay at the frontier, or fund the public-benefit mission, unless customers keep buying tokens fast enough to justify that buildout.
Losses, a confidential prospectus and a delayed listing
The 2025 scorecard in the leaked prospectus is hypergrowth with accounting losses that dwarf revenue. Nearly $4.6 billion of revenue sat against a $42 billion net loss; about $34 billion of that was an accounting charge on financing that could convert into shares, not cash spent running the business, and operating loss excluding those writedowns still exceeded $8 billion. Anthropic held $20.28 billion of cash, cash equivalents and short-term investments at year-end 2025, per the same reporting. Anthropic declined to comment.
May 2026's $965 billion valuation is the last confirmed private mark in these sources. CNBC says some backers have discussed a listing valuation more than double that figure, above $2 trillion, and that the debut is likely after the November 2026 midterms rather than in the original late-September window. A public issuer S-1 was not among the documents opened for this overview; the figures above come from a prospectus Reuters reviewed, not from a confirmed EDGAR posting.
What to watch: Haiku 5.5, the IPO paper and two Pentagon clocks
The next model milestone is Claude Haiku 5.5, which Anthropic still describes only as arriving in the coming weeks after Sonnet 5.5. The next commercial milestone is whether Frontier Academy actually produces credentialed engineers in early 2027, or remains a recruiting brand around the Partner Network.
The IPO clock is the public S-1 and whether a listing after the midterms can digest $518 billion of planned infrastructure, two-customer concentration, and a safety narrative that the prospectus itself, per Reuters, treats as existential. The Pentagon clock is split: an August district-court block on a broader blacklist, and a September 25 D.C. Circuit decision that lets the Department of War keep Claude out of its own systems and defense-contract work. Anthropic's next legal step on that appellate loss was not confirmed in sources opened for this overview.