What are AI benchmarks?

What is it

A benchmark is a standardized test that measures a model's ability on a specific task. It is a fixed set of questions or problems with known answers, a way to score them, and a final number that lets you compare models against each other.

Benchmarks answer narrow questions: "how well does it reason about PhD-level science?" (GPQA), "does it fix real GitHub issues?" (SWE-bench), "which answer do people prefer in conversation?" (LMArena). Each one covers a thin slice of what we loosely call a model's capability.

A benchmark measures what it measures, not "intelligence" in general. A high score on a saturated benchmark tells you very little.

Mental model

A benchmark is like a standardized exam. It is useful for comparing at scale, but it carries the same problems: you can study for the test instead of learning the subject, and once everyone scores 95, the exam stops discriminating.

Three families:

• Knowledge: questions with a single answer. MMLU-Pro, GPQA. Easy to score, easy to contaminate.
• Agentic / task: the model has to achieve something verifiable in an environment. SWE-bench, Terminal-Bench. More expensive to run, closer to real use.
• Human preference: people vote blind between answers from two models. LMArena. Captures "feels better," but it is subjective and gameable.

Two problems you need to know about:

• Contamination: if the benchmark questions, or text derived from them, ended up in the training corpus, the model recalls the answer instead of reasoning it out. Audits find significant contamination across much of the popular benchmark suite.
• Saturation: when frontier models cluster in the 90–94% range, the gap falls within the noise margin. The benchmark no longer separates one model from another.

How it's used

To read a leaderboard without fooling yourself:

1. Look at several types at once: a knowledge test (MMLU-Pro, GPQA), a preference arena (LMArena), and an agentic suite (SWE-bench, Terminal-Bench). The signal is agreement across them; a single number on its own says almost nothing.
2. Check the benchmark's date and whether it is saturated. If the top spots are all above 90, find a harder one.
3. Prefer benchmarks with a private or rotating test set: they resist contamination better.
4. Distrust comparisons where each model reports its own run with its own prompting.
5. For your actual use case: build your own evaluation set with your own data. No public benchmark predicts performance on your specific task.

When to use it / when not to

Public benchmarks are good for:

• Quick elimination: a model that does poorly across the board probably will not work for you.
• Seeing where research is heading, by watching which benchmarks climb fast.
• Roughly comparing model families before investing in your own testing.

They are not good for:

• Deciding which model to ship for your task: your evaluation on your data decides that.
• Comparing models that are all in the saturation zone.
• Drawing conclusions from a single benchmark or a one-to-two-point difference.
• Measuring things no benchmark covers well yet: following long instructions, staying coherent over hour-long tasks, judgment about when to stop.