On October 26, 2026, Anthropic filed its S-1 form with the SEC. 80% of first-half revenue came from token-metered consumption (US$13.1 billion), with subscriptions adding US$3.2 billion. It’s a single data point, from a single company, in the wholesale layer of the market. It proves nothing on its own – but it’s particularly significant in context. AI is struggling to turn a profit, and giants like Microsoft and Salesforce are restructuring their pricing strategy, blending seats with meters.
It’s worth asking: are we at the start of an industry-wide shift in pricing strategy?
Summary
- 80% of Anthropic’s revenue in the first half of 2026 came from users paying for what they consumed (via API), rather than a conventional subscription plan. In 2025, that share was 83%.
- Access to Anthropic’s API is an input for building more software. What matters is whether usage-based pricing trickles down to customer-facing applications, and there are signs it is: Microsoft, Salesforce and support vendors are already combining it with seats.
- What we’re seeing is hybrid pricing, not the end of seat-based pricing: the seat becomes the floor, and consumption becomes the ceiling.
- The same S-1 shows the cost of that model: revenue with no contractual floor against compute commitments worth hundreds of billions.
The data: consumption first, subscriptions as a bridge
The S-1 calls its main model “pay-for-intelligence”: customers pay for usage, “measured in tokens and priced according to the model selected” (p. 139). In 2025, consumption revenue was US$3.796 billion against US$789 million from subscriptions, an 83%-to-17% split. That ratio dipped slightly in 2026, but consumption grew very sharply in absolute terms. The document describes subscriptions as a “bridge” to more consumption once a user exhausts their plan’s capacity.
The company also reports net revenue retention of around 900% as of September 30, 2026. That’s a metric defined by Anthropic that measures only its usage business, so it isn’t comparable to the retention benchmarks of a typical SaaS company.
In April 2026, The Information reported that Anthropic had started charging Claude Enterprise a flat fee of US$20 per user per month, plus token consumption at API rates. Fredrik Filipsson, of Redress Compliance, told The Information that those customers had previously been paying up to US$200 a month per user in exchange for a discounted token allowance, and estimated that costs for heavy users could double or triple. An Anthropic spokesperson told The Register that the transition had begun in November 2025 and that the single US$20 seat launched in February. The Claude help center now explains it this way: the seat covers access only, and every token is billed separately.
Kyle Poyar commented on this in late April in a LinkedIn post and repeated it in his State of B2B Monetization 2026 report: Anthropic “cut the price of Claude Enterprise seats while moving much more aggressively toward usage-based pricing.” The survey behind the report, which polled 230 software companies between April and May 2026, shows the same trend: 37% already use hybrid pricing, combining seats with consumption or other models, up from 25% a year earlier.
But who exactly are these companies charging by usage?
Why one data point isn’t enough: Anthropic sells the input
Charging per token is what you’d expect from a company selling access to an AI model. The real question is whether that meter reaches the applications built on top of it, and the 2026 evidence suggests it does:
Microsoft combines per-user licenses & usage-based billing
On June 1, GitHub Copilot moved chat, agents and code review to AI Credits with usage-based billing, and on June 16 Copilot Cowork launched with per-task pricing, charged on top of the US$30-per-user license. On September 25, Microsoft formalized the split: “everyday” AI stays under the per-user license, while “advanced” AI (long-running agents, Cowork, frontier models) runs on Copilot Credits based on usage.
Salesforce complicates its pricing model
Agentforce is now sold three ways: US$2 per conversation, Flex Credits at US$500 per 100,000 credits, or per-user licenses starting at US$125 a month.
Customer support platforms only charge for what they resolve
Zendesk and Intercom only charge for successful AI resolutions – that is, outcome-based pricing.
What do buyers think of usage-based pricing?
A Futurum survey from the first half of 2026 found that:
- 43% of buyers prefer usage-based pricing models.
- 27% prefer outcome-based pricing.
- Fewer than one in five still prefer the classic per-user pricing model.
On the vendor side, according to Poyar’s survey, 37% already use a hybrid model, and only 5% of investors preferred seat-based pricing.
Our read on the mechanism: when the model provider charges by consumption, the variable cost flows to everyone who resells or packages that model. They can absorb it and lose margin, or pass it on. GitHub’s switch from a flat fee to credits tied to token usage was read in the market as the end of the fiction of selling coding agents the way you’d sell a gym membership.
The seat isn’t disappearing
GitHub kept Copilot’s base prices, and Microsoft kept the US$30 seat. Salesforce still sells per-user licenses. Poyar sums it up this way: most companies are migrating to models that combine seats and consumption, not to pure usage-based pricing.
Usage-based shouldn’t be confused with outcome-based, either. Futurum notes that Microsoft’s scheme gives the buyer a predictable floor for broad deployment, but shifts the problem of forecasting the cost of long-running agents onto FinOps teams. Paying for tokens means paying for activity; paying for resolutions means paying for outcomes. These are different things.
That’s why the careful way to put it is this: the seat goes from being the price to being the floor, and consumption becomes the ceiling. Charging for an outcome or for capacity instead of by the hour is the same logic we examined in our piece on productized services.
Why consumption competes for a different budget
The prospectus defines its agents as AI systems “that work alongside people to carry out complex, multi-step, professional-grade work,” with the permissions, connections and context the organization grants them (p. 8). It presents Cowork as the extension of Claude Code to “knowledge workers in general” (p. 9). The vocabulary is that of labor you hire, not software you buy.
The contrast with seats is our own reading, and it’s shared by Jason Lemkin, of SaaStr, who wrote that this money comes out of headcount and new-project budgets, and that seat-based pricing would have confined it to the software budget.
How consumption reaches the customer
Consumption doesn’t arrive on its own. Revenue from the Amazon, Google and Microsoft marketplaces went from roughly 11% of revenue in 2023 to 32% in 2024, 47% in 2025, and 48% in the first half of 2026. As of June, they accounted for 58% of accounts receivable. For the buyer, that often means the token bill arrives bundled with their cloud bill.
The other channel is services. Ode with Anthropic launched on July 15, 2026, with Blackstone and Hellman & Friedman among its investors, and the S-1 includes it in the company’s go-to-market strategy for mid-sized companies (pp. 10 and 141). The prospectus doesn’t disclose how much of Ode belongs to Anthropic or how much it bills. For agencies and consultancies, implementation is no longer their exclusive turf: the model provider wants to do it too.
Does usage-based pricing fix AI companies’ financial mess?
The usage-based model has a cost, and it’s laid bare in Anthropic’s prospectus. The company posted a net loss of US$80.7 billion in the first half of 2026, of which US$72.4 billion came from changes in the fair value of convertible notes and other instruments, with no cash outflow (p. 64). The operating loss was US$693 million, though the second quarter saw an operating profit of around US$1 billion.
The central problem with usage-based pricing is the timing mismatch. Even if you can charge a base seat, usage-based pricing rejects the revenue stability that’s central to the SaaS model. Revenue is variable and, according to Fortune’s analysis of the S-1 itself, customers can cut their consumption at any time, while the company is locked into compute commitments worth several hundred billion dollars. The exact magnitude depends on what you count:
- Fixed or minimum purchase commitments of about US$265.6 billion as of June 2026, with terms ranging from one to 20 years.
- Plus US$321.4 billion in leases not yet commenced, with some overlap between the two figures.
- And roughly US$245 billion more in agreements signed or amended since June 30, some of them subject to conditions.
On top of that, there’s pricing pressure: the S-1 warns that profitability could suffer if the company is forced to cut token prices, and the cost of running models keeps falling amid intensifying competition. The five largest customers accounted for roughly 25% of revenue for the nine months through September, and the document acknowledges that open-weight models could put pressure on its prices (p. 42).
Put another way, usage-based pricing offers growth with no ceiling – and no floor, either. Whoever charges by usage shares demand risk with the customer.
What this pricing shift means for SaaS, agencies and buyers in Latin America
SaaS
When you charge per seat, the most a provider can bill is the customer’s headcount multiplied by a fixed price, and that money comes out of the software budget. When you charge by consumption, the bill grows as the user weaves AI into its workflows. The bet is that software stops competing within the software budget and instead eats into salaries and services, promising unprecedented productivity gains. Will that actually happen?
Agencies and studios
There are two paths: partner with the labs, or compete on what a consultancy with privileged model access doesn’t have, which is judgment and knowledge of the client’s context. That’s the argument in our guide to choosing a content agency: production is cheap, judgment is expensive.
Buyers in the region
The bill is variable and in dollars, exposed to exchange-rate swings, and is often paid through a pre-existing commitment with AWS, Google Cloud or Azure. Microsoft, for instance, will turn on usage-based billing by default for new Copilot Business licenses sold through CSPs starting December 1, 2026, with a default cap of 4,000 credits per user per month. It’s worth negotiating spending caps and keeping an eye on open-weight alternatives.
If your product is charged per seat and does work an agent can do, your customer already has another way to pay for that work. Anthropic’s S-1 doesn’t prove the shift is universal, but it does show that the model provider already lives off it.