Usage-based pricing
In short
Usage-based pricing is when you pay for how much of a service you actually use, rather than a flat fee. AI developer platforms charge this way, usually per million tokens of text the model reads and writes.
- Also called
- Pay-as-you-go pricing, Consumption-based pricing, Metered billing, Pay per use
- Topic
- The AI business
- Level
- Beginner
- First described
- Pay-for-what-you-use pricing is much older than AI. An early cloud landmark was Amazon S3, launched on 14 March 2006 at $0.15 per gigabyte stored per month, with no minimum fee1.
The bill rises and falls with consumption. Use nothing and you pay little or nothing; use a lot and the bill grows in step. It is how electricity is sold, and how most cloud computing has been sold since Amazon Web Services launched its S3 storage in 2006 with a promise that developers "pay only for what they consume"1.
AI model APIs are the clearest case. Anthropic charges per million tokens (MTok), with separate rates for input and output: Claude Opus 5.5 costs $4 per MTok in and $20 per MTok out, while Claude Haiku 5.5 starts at $0.10 in and $0.50 out2. Discounts also count usage: Anthropic's Batch API halves both rates for jobs that can wait2.
Usage pricing is spreading into workplace tools too. GitHub now measures Copilot chat and agent use in AI credits, where one credit is $0.01 and the cost depends on the model and the number of tokens used3. Salesforce charges for its Agentforce agents per conversation or per action, as well as per user4.
The upside is fairness: light users pay less. The downside is unpredictable bills, which is why vendors add budgets and spending caps. GitHub lets companies set budgets that decide whether extra use is allowed, billed or blocked3.
It is not the same as seat-based pricing, which charges a fixed fee per person whatever they use.
Example
A developer sends a prompt of 2,000 tokens to Claude Opus 5.5 and gets 1,000 tokens back. At $4 per million input tokens and $20 per million output tokens2, that request costs about 2.8 US cents.
Why it matters
It is how almost every AI product is paid for behind the scenes. Heavy use of AI agents can run up large bills quickly, so knowing the meter matters.
Questions people ask
How is AI API usage priced?
Mostly per token, a small chunk of text. Providers set one price for tokens you send in and a higher one for tokens the model writes. Anthropic, for example, lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens2.
Is usage-based pricing cheaper than a subscription?
It depends on how much you use. Light users usually pay less than a flat fee. Heavy users, especially those running AI agents for hours, can pay far more, which is why services such as GitHub Copilot offer budgets and spending caps3.