Evaluation means running a model on a set of test cases and scoring the results. Some answers can be graded exactly, such as a number, a category or code that passes a test. Others need judgment, so they are graded by people, or by another model working from a checklist.
Public benchmarks give a shared way to compare models, but they only cover the tasks their authors chose. A custom evaluation built from your own documents, questions and past failures tells you more about whether a model will work for you.
Results depend on many settings: the prompt, the number of attempts, the reasoning effort, the tools available and the time limit. A score without those details is hard to interpret, and a small test set gives a noisy score.
Define what success means before you test, keep test data separate from the examples used to tune prompts, and rerun the evaluation whenever the model, prompt or tools change.