Anthropic: pay per token, plus subscriptions

Anthropic is a pay-per-token business. According to its S-1, usage is 80% of revenue. Subscriptions are the other 20%, and their real job is to get people using Claude so their companies start paying for tokens.

Anthropic filed its S-1 on October 26, 2026. Kyle Poyar, who writes Growth Unhinged, pointed out on LinkedIn that pricing gets its own risk factor.

This page is my read of how Anthropic makes money, mapped onto the models on this site. The numbers come from the S-1. The diagram is my interpretation, because nobody outside Anthropic can see the whole picture.

The core: you pay for the intelligence you use

Anthropic calls its main revenue model "pay-for-intelligence." Customers pay for usage, metered in tokens and priced by the model they choose. We would call that the Model and inference APIs business model. Pay-for-intelligence is smart positioning compared to pay-for-tokens or usage, because intelligence implies you're getting valuable results and outcomes, even if you're not paying by the outcome specifically.

According to the S-1, consumption revenue grew from $288 million in 2024 (75% of revenue) to $3.8 billion in 2025 (83%) to $13.1 billion in the first half of 2026 (80%). Subscription revenue grew from $98 million to $789 million to $3.2 billion over the same periods. Anthropic expects consumption to stay "the substantial majority" for the foreseeable future.

According to the S-1, the subscription share fell from 25% in 2024 to 17% in 2025, then rose to 20% in the first half of 2026. Usage still dominates. The S-1 doesn't break subscriptions down by product, so it doesn't say what drove the rise.

Subscriptions feed the usage business

Pro and Max look like AI subscriptions: a free tier, then paid plans with bigger usage limits. But the S-1 describes them as a way in, not the end game. Individuals use Claude for their own work, then push for it inside their company. It's a classic bottoms-up sales strategy.

The S-1 also describes a bridge: subscribers who hit the limit in their tier can move up a tier or buy more usage directly, which turns subscription revenue into usage revenue. That's the Subscription for usage credits pattern, where a fee buys a set allowance and heavy users pay for more.

Enterprise works the same way. The Register reported in April that Anthropic stopped bundling tokens into Enterprise seats and now bills usage on top at API rates. According to the S-1, Claude Enterprise users who didn't use Claude Code spent less than $75 a month on average, while the top 5% of Claude Enterprise users who used Claude Code spent about 20 times that.

Anthropic's retention metric points the same way. According to the S-1, Net Revenue Retention leaves out Pro and Max subscribers, because fixed-price plans don't reflect the expansion and contraction the metric is meant to capture.

Enterprise deals look like SaaS but bill like an API

According to the S-1, a growing number of enterprise customers sign non-cancelable annual or multi-year contracts with a minimum spend, in return for a committed price list. That sounds like SaaS. But the S-1 says these contracts produce mostly consumption revenue. The customer commits to a floor, then pays for what they use.

Many companies buy Claude through the Amazon, Google or Microsoft cloud marketplace, on the cloud account they already pay for. According to the S-1, that route carried 48% of revenue in the first half of 2026, up from 11% in 2023. The cloud provider sends the bill, collects the money and keeps a fee. Anthropic still sets the price and owns the relationship with the customer.

According to a Reuters estimate reported by Coinpaper, that fee came to about 16 cents of every marketplace dollar in 2025. According to Latent Space, Anthropic books those fees as a sales cost, so they don't show up in its gross margin.

According to the S-1, Anthropic had more than 11,900 customers paying $100,000 or more a year as of September 30, 2026, up from about 20 at the end of 2023.

According to the S-1, Net Revenue Retention was about 900% as of September 30, 2026. That means customers who were paying a year earlier now spend about nine times as much. The S-1 measures it by comparing the latest three months with the same three months a year before, leaving out Pro and Max. It also warns that recent growth partly reflects early customers expanding fast, which may not last.

ARR becomes "Run-Rate Revenue"

The S-1 doesn't report ARR. It reports Run-Rate Revenue: usage fees from the last 28 days times 13, plus this month's subscription fees times 12. (Thirteen 28-day periods make a year.)

A 28-day window makes the number move with the calendar. The S-1 says total customer spend fell about 20% in the two weeks ending January 2, 2026, compared with the two weeks before. Measure Run-Rate Revenue over the holidays and it drops. Measure it the week after a big model launch and it jumps.

According to the S-1, Run-Rate Revenue also counts spending from customers Anthropic later cut off for breaking its usage policy, so part of the number won't come back. And in this version of the filing, the Run-Rate Revenue figure itself is left blank. The S-1 explains how the number is calculated, but not what it is.

That makes it different from ARR at a SaaS company. ARR is usually backed by contracts, where customers have agreed to pay that amount over the next year. Run-Rate Revenue is last month's usage, scaled up to a year. Nobody has promised to keep spending at that rate.

Costs don't move with usage

The Model and inference APIs diagram says cost is GPU cost per token times tokens. Anthropic's S-1 describes something different: most compute is bought under multi-year contracts, 12 to 18 months before it's available. According to the S-1, it's shared across research, model development, serving customers and internal use, and it's mostly fixed in the short term.

So the margin question isn't "what does a token cost to serve?" It's "is revenue growing faster than the compute we already committed to?" According to the S-1, compute and infrastructure fell from 201% of revenue in the first half of 2025 to 55% in the first half of 2026.

Which means in the first half of 2025, Anthropic was spending about $2 on compute for every $1 of revenue it brought in. Now that's not the case: a year later it was about 55 cents. That compute also trains new models, so it's more than the cost of serving today's customers.

According to the S-1, Opus 5.5 cut customers' cost per task by about 40% compared with Opus 5. The S-1 also says quarter-over-quarter revenue growth went from 48% to 132% after Opus 4.5 launched at a lower price, and it credits several factors, lower prices among them. The S-1 says customers have generally expanded their use after launches that lower the cost per task.

On 20VC x SaaStr, Rory O'Driscoll points to the S-1's figure that over 80% of revenue comes from models less than six months old. His read is that sellers of per-token APIs have to keep paying for costly new models to keep earning.

The S-1 names agents as a risk: it says agentic products are "structurally more compute-intensive," and if pricing doesn't keep up, compute efficiency gets worse as agent usage grows.

Pricing is a work in progress

The pricing risk factor says Anthropic has changed its pricing models before and expects to again. It names lower prices, shorter contracts and "outcome or task-based pricing" as possibilities.

This year already shows it. As The Register reported in April, Enterprise moved to seats plus usage. And The Register reported that Anthropic briefly tested making Claude Code available only on Max for about 2% of new signups, then backed off after the backlash. Kyle Poyar's point is that pricing models used to last 18 to 24 months, and now one can be out of date in six.

Outcome or task-based pricing would move Anthropic toward Outcome-priced AI agents or Pay per job. Kyle Poyar notes that OpenAI has started testing outcome-based pricing with some enterprise customers, and that Cognition, the company behind the Devin AI coding agent, now offers a productivity guarantee. He asks whether customers want to pay for intelligence, or for the work it delivers.