Glossary · Model mechanics
Benchmark
Also said as
benchmarksmodel benchmarkleaderboard
A benchmark is a standardized test that measures what an AI model can do — accuracy, refusal rates, hallucination rates — so models can be compared.
Why it matters to your business
Vendor marketing says a model is great; benchmarks are where the numbers come from — like a 47% failure rate on factual questions that never made it into an ad.
The catch · what vendors don't say
Scores depend entirely on the task tested — the same model can score brilliantly on one benchmark and poorly on another, and 'newer' often measures worse.
