Evals (evaluations) are the tests you use to measure whether an AI model, or a system built on one, does your task well. They have two parts: a set of cases, each with the expected answer or a criterion, and a way to score them. You run them again every time you change model, prompt or provider.
A question from the app. Pick an answer.
You moved to a cheaper model and the bill dropped. Now what?
Saving can make results worse without it showing in the invoice.
With your cases you see if the drop was worth it or came at the cost of worse answers.
The cap controls the total, it does not measure quality. They are two different things.
It depends on the task. If the AI sorts complaints into "shipping", "payment" or "product", comparing against the expected answer is enough. If it has to return data, you check rules: the JSON is valid, it includes the order number, it stays under 50 words.
For open answers, like a summary, you use another AI as a judge with a list of criteria, or a person who reads and grades. The automatic judge is fast and cheap, but it also makes mistakes: check part of it by hand.
In programming, tests run every time you change the code. Evals play that role when you change model, prompt or provider. Example: you switch to a cheaper model and the bill drops. You run the evals and see whether hits stay at 92 out of 100 or fell to 80. Only then do you know if you really saved.
One difference: the same model can answer the same question differently. That is why the result is usually a percentage rather than a "pass or fail", and sometimes each case is run more than once.
A benchmark is a fixed exam given the same way to several models to compare them. Evals are the broader practice of measuring whether your system (model, prompt, data and tools) works for what you need. A benchmark can be one of the pieces of an evaluation, along with your cases and the way you score them.
"I tried it with three questions and it worked." A few easy cases say nothing: add the hard ones and the ones that failed before. "The AI judge is objective." It has its own biases; compare it with a person. "I already evaluated it once." If the model, the prompt or the data change, you have to run them again.