Anthropic Company Profile: Revenue, Business Model & Strategy

What Does Anthropic Do?

Anthropic trains frontier language models and sells them together with the agent software built on top of them. The product line in September 2026 looks like this:

• Claude Opus 5.5: the flagship model, released September 22, 2026 as the first of the Claude 5.5 family. It leads Anthropic's own benchmark table on SWE-bench Pro (89.9 percent) and Terminal-Bench 4.0 (66.4 percent), costs $4 per million input tokens and $20 per million output tokens, and by the company's measurement runs about 40 percent cheaper than Opus 5 on typical work.
• Claude Sonnet 5.5: released September 28, 2026 at roughly half the cost of Opus 5.5, and the default model on the Free and Pro plans.
• Claude Fable and Claude Mythos: the restricted top tier. Mythos was announced on April 7, 2026 and never broadly released; access to Fable 5 and Mythos 5 was suspended in June 2026 under US export controls and restored on July 1, 2026 for approved organizations and vetted users.
• Claude Code: a command line agent for software work, generally available since May 2025 and used at companies including Netflix, Spotify, KPMG and Salesforce. Its run-rate revenue passed $1 billion in December 2025 and $2.5 billion by February 2026.
• Cowork: launched in January 2026, it brings the same execution model to non-technical work through a graphical interface, with open-source plugins for roles such as sales, legal and finance.
• Claude Developer Platform: the token-billed API, with prompt caching, web search, the Model Context Protocol connector and the Claude Agent SDK.

How Does Anthropic Make Money?

Four revenue lines carry the business:

1. API consumption. Developers and enterprises pay per million tokens. Cache reads cost $0.20 per million on Opus 5.5 against $4 for fresh input, which is what makes long agent loops over large codebases affordable.
2. Agent products. Claude Code and Cowork are sold to individuals, teams and enterprises. Enterprise accounts supplied more than half of Claude Code revenue by February 2026, and business subscriptions to it quadrupled in the first weeks of that year.
3. Seat subscriptions. Claude Pro, Max, Team and Enterprise plans, with single sign-on, role based access control and audit logs at the enterprise tier, available to HIPAA-covered organizations since January 2026.
4. Cloud marketplaces. Claude is the only frontier model sold on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Azure Foundry, so customers can buy it with committed cloud spend and run it inside their own environments.

Revenue is concentrated. The draft IPO prospectus shows the two largest customers each accounting for about 12 percent of 2025 revenue, while more than 500 customers spent over $1 million on an annualized basis as of February 2026 and eight of the Fortune 10 were customers.

The Numbers Behind the Growth

Anthropic confidentially submitted a draft S-1 to the Securities and Exchange Commission on June 1, 2026. Reuters reported the contents on September 28, 2026, which put audited figures in public view for the first time:

• Revenue of $4.59 billion in 2025, up 1,088 percent from $386 million in 2024.
• An operating loss of $8.06 billion, widened from $2.98 billion, on total operating expenses of $12.65 billion.
• Compute and infrastructure spending of $7.33 billion, roughly triple the 2024 level.
• A GAAP net loss of $41.97 billion, of which about $34 billion was a non-cash charge on convertible financing raised from Google and Amazon rather than cash burn.
• $20.28 billion of cash, cash equivalents and short-term investments at the end of 2025, and about $518 billion of long-term cloud and infrastructure commitments.

Run-rate revenue moved far faster than the annual accounts suggest: $14 billion in February 2026, above $47 billion in May, and about $65 billion in July. Second-quarter 2026 revenue of $11.5 billion exceeded OpenAI's $6.7 billion for the same period, the first quarter in which Anthropic was the larger business, and came with its first positive adjusted operating income.

Origins: The Split From OpenAI

Anthropic began with a disagreement inside OpenAI in late 2020. Dario Amodei, then Vice President of Research and the leader of the teams that built GPT-2 and GPT-3, and his sister Daniela Amodei, Vice President of Safety and Policy, concluded that the Microsoft partnership was pushing commercial speed ahead of alignment work. They left with five colleagues, among them Jared Kaplan, Sam McCandlish, Tom Brown, Chris Olah, Ben Mann and Jack Clark, and incorporated Anthropic on January 26, 2021.

The structure was part of the argument. Anthropic is a Delaware public benefit corporation, and an independent Long-Term Benefit Trust, whose trustees hold no financial stake, has the right to appoint and replace part of the board. The company raised $124 million in May 2021 and $580 million in April 2022, a round that included $500 million from FTX. It finished training the first version of Claude in the summer of 2022 and held it back for months of internal testing, releasing it only in March 2023, after ChatGPT had already defined the category.

Constitutional AI

The method Anthropic is best known for is Constitutional AI, published in December 2022. Conventional alignment relies on reinforcement learning from human feedback, where contractors rank model answers. That approach is expensive, inconsistent across specialist subject matter, and prone to teaching models to flatter the rater.

Constitutional AI replaces most of that labor with an explicit written constitution, drawn from sources including the Universal Declaration of Human Rights, against which the model critiques and revises its own answers. Claude 2's constitution included instructions derived from the declaration, such as choosing the response that best supports freedom, equality and a sense of brotherhood. The result is a training loop that scales without a large human rating workforce, and a safety story enterprise buyers can read.

Compute: Three Clouds and Several Chip Suppliers

Training frontier models requires power, memory and accelerators at industrial scale, and Anthropic has deliberately spread its supply. Amazon has invested $8 billion, remains the primary cloud and training partner, works with Anthropic through its Annapurna Labs chip division on Trainium silicon, and agreed in May 2026 to supply up to five gigawatts of new capacity. Google invested $500 million in October 2023 with $1.5 billion committed over time, added $1 billion in March 2025, agreed in October 2025 to supply up to one million TPUs, and in 2026 signed a further deal with Google and Broadcom for gigawatts of next-generation TPU capacity from 2027.

In November 2025 NVIDIA committed up to $10 billion and Microsoft up to $5 billion, while Anthropic agreed to buy about $30 billion of Azure compute and to contract up to a gigawatt of NVIDIA systems. AMD followed in July 2026 with 2 gigawatts of Instinct MI450 GPUs and an investment of up to $5 billion. Anthropic also signed a roughly $45 billion lease with Nscale for about 460 megawatts in West Virginia, expected to open in late 2027, and a 20-year TeraWulf data center lease in Kentucky. Micron, Samsung and SK hynix supply memory. The total of these obligations, about $518 billion, is the single largest number in the prospectus.

Interpretability: Looking Inside Claude

Anthropic runs one of the larger research programs in mechanistic interpretability, the attempt to explain what a neural network is doing internally rather than judging it by its outputs. In 2024 the team used dictionary learning, a sparse autoencoder technique, to extract millions of interpretable features from Claude 3 Sonnet, including a feature that activates on the Golden Gate Bridge. Research published in 2025 found that multilingual models process meaning in a shared conceptual space before rendering it in a language, and that Claude selects candidate rhyming words before writing a line of poetry.

That work feeds release decisions. Every frontier model now goes through an automated behavioural audit across thousands of simulated scenarios; Opus 5.5 scored best of any Anthropic model on it and was tested by outside evaluators including METR before release. In September 2026 Anthropic went further and gave independent evaluators permanent, employee-like access to its models, an arrangement Dario Amodei compared to letting a food inspector check the kitchen.

Where the Friction Is

Growth has come with legal and political cost. A 2024 class action established that Anthropic had built an early training library from more than seven million pirated books; a judge held that training on lawfully obtained copies was fair use but that the pirated library was not, and Anthropic agreed in September 2025 to pay $1.5 billion, about $3,000 per book. Final approval came on July 20, 2026, the largest known settlement in a US copyright case. Court filings unsealed in January 2026 revealed Project Panama, an internal effort to buy used books and scan them destructively for training data, which the court treated as fair use because the copies were purchased. Music publishers sued over lyrics in 2023 and again in August 2026 over allegedly torrented compositions, and Reddit sued in June 2025.

The sharpest conflict was with the US government. In February 2026 the Department of Defense demanded that Anthropic drop contractual limits barring mass domestic surveillance and fully autonomous weapons. Anthropic refused, the Pentagon designated it a supply chain risk in March, and a federal judge granted a preliminary injunction later that month, writing that the action appeared to be First Amendment retaliation. Anthropic has also accused DeepSeek, Moonshot AI, MiniMax and Alibaba Cloud of distilling Claude outputs to train competing models.