Anthropic does not publish a full breakdown of its finances, but the shape of its business is clear from its products. The company makes money three main ways: subscriptions from people who pay for Claude, usage fees from developers who rent its models by the token, and contracts from companies and governments that deploy Claude at scale.
Each stream behaves differently. Subscriptions are predictable but capped by how much one person will pay per month. API revenue grows with usage, which means it grows with customers’ success — and with Anthropic’s own computing bill.
Subscriptions: Free, Pro, Max and Team
The consumer business starts with a free tier designed to get Claude into people’s daily habits. Paid plans begin with Pro at $20 a month, aimed at individuals who hit free limits. Above that, the Max plans — $100 a month for five times Pro’s usage, $200 for twenty times — exist for heavy users, especially people running Claude Code for hours a day.
For organisations, Team plans charge per seat, starting at $25 per seat monthly for the standard tier, with a premium seat for engineering-heavy teams. Enterprise agreements are negotiated individually, bundling security controls, audit logs and support.
Subscription revenue is the steadiest money Anthropic makes. It is also the most contested: every rival assistant, from ChatGPT to Gemini, is fighting for the same $20 a month.
The API: renting Claude by the token
The second engine is the Anthropic API. Developers send text to Claude and pay per million tokens processed, with prices varying by model: small Haiku models cost a fraction of the flagship tier, and techniques like prompt caching cut the price of repeated context.
The API reaches customers through three doors. Developers can buy directly from Anthropic’s platform, or through Amazon Bedrock and Google Cloud’s Vertex AI, where Claude sits on the same bill as the customer’s other cloud services. Marketplace distribution sacrifices some margin for reach into large companies that standardise on one cloud.
Inside this stream sits the standout: Claude Code, the company’s agentic coding tool, is sold partly as a subscription feature and partly as metered usage. It passed a $2.5 billion revenue run-rate in February 2026, a figure Anthropic itself highlighted, and it explains why the Max tiers are priced the way they are — serious coding agents consume serious compute.
Enterprise and government deals
The third stream is contracts. Banks, insurers, law firms and public bodies buy Claude with the extras regulated industries require: single sign-on, data-retention controls, compliance certifications and commitments that their data will not train future models.
These deals are fewer and larger, and they lean on Anthropic’s safety reputation. Published evaluations and a written scaling policy give a risk committee something concrete to approve, which is worth money when the alternative is an unexplained model.
Why the mix matters
The balance between the three streams shapes Anthropic’s decisions. Subscription revenue is stable but grows one user at a time and is brutally competitive. API revenue can grow very fast — a single successful product built on Claude can multiply its token consumption overnight — but it is equally quick to shrink if that customer switches models or negotiates harder.
Cloud marketplaces complicate the picture further. Selling through Amazon Bedrock or Google Cloud puts Claude in front of huge enterprise buyers, but the cloud provider takes its cut and owns the customer relationship. Anthropic trades margin for distribution every time a deal closes that way, which is why its direct enterprise sales team matters more each year.
The free tier belongs in this accounting too. Every free user costs compute and returns no revenue; the tier exists as a funnel, converting a fraction of users to paid plans and keeping Claude’s habits-forming presence in front of people who might otherwise standardise on a rival. It is marketing spend paid in GPU hours.
Finally, scale itself is a strategy. Frontier training runs are so expensive that revenue growth is partly defensive: the lab that earns more can afford the next generation of compute, which produces the models that earn more. Anthropic’s business model and its research ambitions are the same flywheel.
The cost side of the ledger
Revenue is only half the picture. Frontier models consume enormous computing capacity, much of it bought from the same cloud giants that distribute Claude, and top researchers command salaries that resemble sports contracts.
Legal exposure is another real cost. Anthropic’s settlement in the Bartz book-piracy case — $1.5 billion, with final approval in July 2026 — shows how training data decisions made years earlier can return as nine-figure line items.
So the business, in one line: subscriptions provide the base, the API provides the growth, enterprise contracts provide the margin story — and compute, talent and lawyers consume a formidable share of all three.
0 Comments