What is Evals?

What they are

Evals, short for evaluations, are repeatable tests for an AI system. You collect a set of example inputs, decide what a good output looks like, and score the system against them. Run the same set again after you change the model, the prompt, or the settings, and you can see whether things got better or worse.

A public AI benchmark is one kind of eval, built to compare many models on a general skill. Evals in the everyday sense are usually your own, built around your actual work.

How they work

A basic eval has three parts:

• Cases. Real examples of the task: support questions, documents to summarize, bugs to fix. Include easy ones, hard ones, and the failures you have already seen.
• A grader. Some checks are exact, like "does the code pass its tests" or "does the answer match." Others are fuzzy, so a person scores them or another AI model acts as the judge against a written rubric.
• A score you can track. Pass rate, average rating, cost, and speed, recorded each time so you can compare runs.

Because model output varies from run to run, good evals run each case several times and look at the average.

Why they matter

Public leaderboards tell you how a model did on someone else's tasks, often under conditions the lab chose. They cannot tell you whether a model handles your customers, your documents, or your codebase. Evals can. They are how you decide between models, notice that a provider's update changed behavior, and confirm that a cheaper model is good enough for a routine job.

Without evals, "it seems better" is a guess. With them, a prompt tweak or a model swap is a measured change. They are also the way to catch regressions, where a fix for one case quietly breaks another.

Pitfalls

Evals can mislead. If your cases are too few or too easy, every model looks great. If the same cases are used to tune a prompt and to judge it, the score overstates real performance. An AI judge can favor longer or friendlier answers. A system can also learn to satisfy the grader without doing the task, which is reward hacking.

Start small. Twenty real cases you actually care about, scored consistently, beat a thousand generic ones. Keep adding the failures you meet in use. Pair the numbers with reading actual outputs, since scores hide the way a system fails. Evals also help you judge AI agents, where the question is whether the whole task got finished, not whether one answer was good.