Calling an AI model is no longer a research project; it is a line item in a budget. Every request is metered in tokens, priced per million, and shaped by a small set of variables: model size, speed, and context window. Understanding those numbers helps teams spend wisely.
The basic unit of cost is the , a short piece of text. A typical page of English is a few hundred tokens. A long chat thread can reach tens of thousands. If a provider charges three dollars per million input tokens, sending a 2,000- prompt costs about six tenths of a cent. Multiply that by thousands of users, and the bill grows fast. Output tokens are usually priced higher than input tokens because generating them uses more compute.
Three levers control the bill. The first is model choice: a small, fast model handles most tasks at a fraction of the price of the largest flagship. The second is prompt design: tight context, clear instructions, and of repeated prefixes all reduce waste. The third is routing: a smart layer sends simple requests to cheap models and only escalates hard ones to premium endpoints.
Below the public price, providers wrestle with the real cost: GPUs, energy, and skilled staff. Healthy comes from running models at high utilization across many customers. For buyers, the takeaway is simple: measure per-feature cost, design prompts to be lean, and stay alert as and alternatives keep shifting under your feet.