The headline making rounds last week: Anthropic reached an annualised run rate of $30 billion, while OpenAI is reported at $24 billion. The thing worth looking at isn’t the number — it’s the shape of the number.
Most of Anthropic’s enterprise revenue comes from agentic workflows, not from people typing into a chat box. In our shop and across the half-dozen client projects I’ve sat in on this quarter, that’s exactly what we’re seeing.
What changed
A year ago, “enterprise GenAI” meant a fancy chat interface bolted onto a company’s data, with humans driving the conversation. The economics of that pattern were limited: per-seat licensing, weekly active users, the kind of thing that scales with headcount.
In 2026 the dominant pattern is different. The model is invoked by systems, not by humans:
- A document review pipeline that processes 50,000 contracts a night.
- An agent that runs hourly, queries 30 internal services, and fills out a structured report.
- A data validation step in a pipeline that touches every record before it reaches the warehouse.
- A codegen workflow that opens dozens of pull-requests a day on internal repos.
Each of these can burn more tokens in an hour than a thousand human chat users do in a week. Once you wire even one of these into a real production system, the per-user-seat model stops being the unit of measurement.
Why Anthropic specifically
Three things, from the conversations I’ve had with peers building on it:
One. Claude’s tool-use reliability — both first-party and via MCP — has been ahead of competitors on the boring axis of not breaking when called 50,000 times in a row.
Two. Their pricing for enterprise tiers has been more predictable than the alternatives. Predictability matters more than headline price when you’re forecasting a workload.
Three. MCP itself was Anthropic’s bet, and it paid off. Customers building on Claude get the easiest path into the connector ecosystem.
This isn’t a “Claude wins” post — the numbers will move, OpenAI’s enterprise push is real, and Google’s agent stack is closing fast. But the shape of the GenAI market in 2026 is settled: it’s an infrastructure market with agentic workloads on the demand side.
Sources: