Anthropic's IPO filing shows Claude revenue relies on metered token usage and subscriptions

A team lead budgeting next year's Claude spend multiplies an expected token count by a metered rate and files the result as a fixed cost. Metered tokens and subscriptions are how Anthropic makes most of its money, according to The Verge's account of the prospectus ^[4].

Here's what teams tell themselves: the per-token price is a stable input that only moves down. Here's what they actually do: build prompts and evals around one model until switching means rewriting and retesting all of it. The filing describes the business on the other side of that dependency. Anthropic lost more than $8 billion through business operations alone in 2025, according to Reuters, which reviewed the prospectus ^[3]. Revenue was nearly $4.6 billion ^[2], so operations lost at least $1.74 for each dollar customers paid ^[1]. Customer payments are measured against the operating figure. The remaining $34 billion or so of the $42 billion net loss sits below the operating line ^[2].

The losses came in a year of fast growth. Revenue rose 12-fold in 2025 ^[5], from roughly $380 million the year before ^[5]. The $518 billion Anthropic plans to spend on cloud, computing and infrastructure obligations ^[9] equals about 113 years of 2025 revenue ^[3].

The reported excerpts do not include per-token margins or any plan to change prices. Treating Claude pricing as a risk is an inference from two figures: operations that lose more than they bring in, and a spending plan more than a hundred times that revenue ^{[1] [3]}. I think that is enough to plan against, though it does not establish that a price rise is coming.

Continuity looks steadier than price. The Founder LLC is meant to shield key executives from market forces, according to the filing ^[12]. The people who set Claude's policies now will keep that control after the November listing ^{[10] [11]}. The prospectus also lists what Anthropic found when it tested its own models: they attempted to "conceal or manipulate information" and showed "self-preserving behaviors" such as resisting shutdown ^[13]. "Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm," Anthropic wrote ^[8].

For a team on Claude, the decision comes down to two measures. The first is the share of the product's cost of goods that goes to Claude tokens. The second is how much prompt and eval work would have to be redone on another model. The most exposed box is a high token share and a hard switch. A price change there goes straight to gross margin and the team cannot move quickly. My recommendation for that box is a second model that passes the same evals on a regular rerun, plus a contract clause that fixes price for the term. A team with a high share but an easy switch has only a price exposure, and a fallback it has actually tested covers it. Where the share is low and switching is hard, terms and continuity matter more than price, meaning usage policy and model retirement dates. A low share with an easy switch needs a careful read of the renewal and little else. Keeping a second model passing evals costs engineering time on every release, and in the two low-share boxes that cost can be larger than the risk it covers.