Anthropic sells Claude through three commercial models, and most buyers only understand the one they happen to be on. There is per seat pricing for the application products, per token pricing on the API, and dedicated capacity for the largest and most latency sensitive workloads. Each one prices on a different basis, each one bends in a different place, and the right deal usually involves more than one of them at once. If you are sizing a Claude purchase in 2026, the first job is to understand which model you are buying through and what actually moves the number in each. This guide lays out all three from the buyer side.
The seat based products, Claude for Work in its Team and Enterprise forms, price on a per user per month basis. You pay for the people who have access, not for what they consume, which makes the cost predictable and easy to budget but also easy to overpay. The two questions that move a seat deal are how many seats you actually need and what the per seat rate is, and on enterprise agreements both are negotiable in ways the published figures do not advertise.
Seat count is where most of the waste hides. Organizations buy seats for everyone who might use Claude, then discover that a fraction of those people use it regularly. You are paying full price for licenses that sit idle. The discipline is to right size the seat count to real usage, which means measuring active users before you commit rather than buying for the org chart. The per seat rate then moves with volume, term length, and the Enterprise feature tier you select, including the larger context window and the admin and security controls that the Enterprise plan carries over Team. A buyer who walks in with usage data negotiates a smaller, cheaper seat block than one who buys for headcount.
The API prices on consumption: you pay per input token and per output token, at rates that differ by model. Opus is the premium tier, Sonnet sits in the middle, and Haiku is the economical option, and the spread between them is large. Output tokens bill at roughly five times the input rate across the models, which is why output discipline matters so much to an API bill. This is the model that powers most production applications, and it is the one where consumption can climb fastest, because nothing caps it the way a seat count caps the seat model.
Three structural levers move an API bill before you ever talk price. Model routing, sending each request to the cheapest model that handles it well rather than running everything on Opus, typically cuts aggregate spend by forty to seventy percent on a mixed workload. Prompt caching cuts the cost of repeated stable context by up to ninety percent. Batch processing, for work that does not need an immediate answer, runs at roughly half the standard rate. These are engineering levers, not negotiation levers, but they set the run rate that everything else is priced against, which is why the technical and commercial sides of an API deal cannot be separated.
On top of the levers sits the commercial structure. At enterprise scale, the API is sold through committed spend agreements: you commit to a level of spend over a term in exchange for a discount off the standard rates. The commitment bands, the discount that comes with each, the rate on overage above the commitment, and the treatment of unused commitment are all negotiable, and they are where a buyer side advisor earns the fee. The list rates are the starting point, not the deal.
The third model is dedicated capacity, where Anthropic reserves throughput for you rather than serving you from the shared pool. This is for the largest workloads and for applications where latency consistency matters enough to pay for guaranteed capacity. It prices differently from per token consumption, closer to a reserved infrastructure model, and it is the least transparent of the three because it is always sold through a direct conversation rather than a published rate.
Dedicated capacity makes sense for a narrower set of buyers than the other two. If your workload is large, steady, and latency sensitive, reserved capacity can be both cheaper per unit and more predictable than buying the same volume through standard consumption. If your workload is spiky or still ramping, committing to reserved capacity you cannot keep full is an expensive mistake, because you pay for the reservation whether you use it or not. The decision turns on how steady and how large your demand really is, which again comes back to having real consumption data before you commit to anything.
Large enterprises rarely buy through a single model. A typical arrangement bundles seats for the people using the Claude applications, a committed API spend for the production workloads, and possibly dedicated capacity for a latency critical service. Anthropic is happy to package these together, and the bundle is where both opportunity and risk live. The opportunity is that a larger combined commitment earns a better discount across the whole relationship. The risk is that bundling obscures the individual prices, so you cannot tell whether the seat rate, the token rate, and the capacity rate are each competitive, or whether a good headline discount is hiding a poor component price.
The buyer side discipline on a bundle is to price every component on its own before accepting the package, so you know what each piece costs and can negotiate each on its merits. A bundle should earn you a better price on every component than you would get buying them separately, not just a better price on the easiest one to benchmark. That requires knowing what comparable enterprises pay on each model, which is exactly the kind of benchmark that is hard to assemble from the outside and central to what we do.
Claude prices three ways in 2026: per seat for the application products, per token on the API, and as dedicated capacity for the largest steady workloads. Seat deals bend on count and rate, API deals bend on the engineering levers and then on the committed spend structure, and dedicated capacity bends on how steady your real demand is. Most enterprise deals combine more than one, and the job is to price each component honestly rather than accept a bundled number you cannot decompose. Our pricing guide lays out the bands, the levers, and the benchmarks in detail. Download the playbook and bring real numbers to the table.
Download the pricing playbook for the bands, the levers, and the benchmarks across seats, API, and dedicated capacity.
Get the playbookWeekly intelligence on Anthropic pricing moves and the buyer side counters that work.