Enterprise AI software is priced on the unit the vendor can meter. A copilot bolted to an existing suite is usually a seat. A model API is usually a token. A platform that wraps both will put a seat, a consumption line, and a commit on the same order. The discount on one line says nothing about the other two.
The three units#
Seat. A named user, a bundle add-on, or a platform fee per employee. The bill is headcount times rate, plus the renewal increase on that rate. Per-seat pricing is the general mechanic. Unused seats are shelfware.
Token. Input and output are different rates. The per-token pricing page states a 2026 frontier range of about $2 to $15 per million input tokens and $8 to $75 per million output tokens, with output on the order of 3 to 5 times the input rate. Cached input and batch inference are cheaper paths on that same page. A prompt that is long, and an answer that is long, are both spend, and the answer is the expensive one.
Commit. Reserved throughput, a prepaid credit pool, or a cloud commit that the AI product burns. The per-token page describes committed throughput discounts of about 20 to 40 percent off pay-as-you-go for the contracts it covers, with deeper tiers at high reserved volume. A commit you do not consume is the same leftover as an oversized EDP. Marketplace spend and cloud commits is how that leftover interacts with AWS, Azure, and Google Cloud.
What to compare#
Compare seat deals to seat deals, and token deals to token deals, at a similar volume. A percent off a seat list price and a percent off a token list price are not one discount. How to read a software discount is the reading order.
Then separate the renewal. Seat fees take a renewal uplift. Token rates move when the vendor changes the public price card, unless the contract freezes the rate. Read the freeze and the commit term as a pair.
The AI and machine learning category collects the vendor pages. Consumption-based pricing is the broader pattern token billing belongs to. Databricks capacity, which is a commit of its own, sits in commitment structures.
Frequently asked questions
- How is enterprise AI software priced?
- Three units show up on the order: a seat or subscription fee, a consumption fee per token or per request, and a commit that prepays capacity in exchange for a lower rate. Many contracts use more than one.
- Why do output tokens cost more than input tokens?
- The per-token page explains output as sequential generation and input as a single pass, which is why output rates run several times the input rate.
- What is committed throughput?
- Committed throughput is reserved capacity at an API, prepaid, in exchange for a lower rate than pay-as-you-go. The per-token page describes that discount, for the contracts it covers, as on the order of 20 to 40 percent off the public rate.