What does an AI benchmark measure?
A benchmark measures one capability under one fixed protocol. It pins down three things: the task set, how the model is run (prompt format, number of attempts, reasoning settings, tools allowed) and the scoring rule, usually accuracy or a pass rate. Change any of the three and the number stops being comparable. A score quoted without its protocol is a claim, not a measurement. Language model benchmarks now range from multiple-choice exams to agent runs inside a real codebase, where the model edits files and a test suite grades the result.
Which AI benchmarks do model comparisons cite?
These public benchmarks recur in model comparisons, and each tests a different skill.
| Benchmark | What it tests | How it is graded | Size |
|---|---|---|---|
| MMLU | Knowledge across subjects from math to law | Multiple choice | 57 tasks |
| GPQA | Graduate-level biology, physics and chemistry | Multiple choice, written by domain experts | 448 questions |
| Humanity’s Last Exam | Expert-level questions across dozens of subjects | Multiple choice and short answer | 2,500 questions |
| HumanEval | Writing Python functions from docstrings | Unit tests | 164 problems |
| SWE-bench | Resolving real GitHub issues | The repository’s tests | 2,294 tasks; Verified subset 500 |
| Chatbot Arena | Preference in open conversation | Crowdsourced pairwise votes | Rolling |
Sizes come from each paper or the official site (read 16 November 2026). Coding comparisons lean on HumanEval and SWE-bench, knowledge comparisons on MMLU, GPQA and Humanity’s Last Exam. Pick the row whose task looks like your work, because an overall rank averages away the row you care about.
Why do AI benchmark scores mislead?
Scores mislead through saturation, leaked test data, uneven settings and the harness around the model. Benchmarks saturate: the Humanity’s Last Exam authors note that models “now achieve over 90% accuracy on popular benchmarks like MMLU” (Phan et al., read 16 November 2026), so the top of the table no longer separates them. Public questions leak into training data, the problem called benchmark contamination. A vendor may run its own model at its highest settings and quote competitors’ self-reported numbers, as the next section shows. And the code around the model counts: LangChain moved its coding agent from 52.8 to 66.5 on Terminal Bench 2.0 while keeping the model fixed (LangChain, 17 February 2026, read 16 November 2026), a result we cover in harness engineering. A benchmark score measures the model plus its harness and settings, not the model alone.
How do we read a vendor’s benchmark table?
We read the methodology note before the headline. When Google launched Gemini 4 Argon, its evaluation note said results for non-Gemini models were “sourced from providers’ self reported numbers” unless stated otherwise (Gemini 4 Argon model evaluation, read 16 November 2026), so the rows came from different runs under different settings; our post on what you can call today walks through what that changes. We then compare only the rows our workload resembles. A vendor’s lead on one evaluation is a reason to test, not a reason to switch: the next step is your own AI agent evals on a private eval dataset built from real cases. When we review an evaluation setup, we expect any LLM judge to come from a different model family than the model it grades. A benchmark tells us what to test first, never what to ship.