What is an LLM benchmark?

A benchmark is a standardised set of tasks with known answers. Run a model against it, score it, compare. Every launch announcement is built from them.

The best known are exam-style: MMLU, introduced in 2020, tests knowledge across 57 subjects. HumanEval, from 2021, gives coding problems and runs the tests.

They exist for a good reason. Before benchmarks, "this model is better" was an opinion. They made progress measurable, and that mattered.

Why the numbers keep disappointing you

You read that a model scores 90% on a reasoning benchmark. You try it. It cannot do your task. You conclude the benchmark is fake.

It is not fake. It measures something. Just probably not what you needed.

Contamination. Benchmarks are published on the internet. Models are trained on the internet. When a model has seen the test during training, the score measures recall, not capability — and nobody can fully verify what was in the training data. This is the central, unsolved problem with benchmark scores.

Saturation. Once everyone scores 90%+, the benchmark stops discriminating. The remaining 10% is often ambiguous or mislabelled questions, so the difference between 92% and 94% may be noise, or skill at guessing badly-worded items.

Undisclosed conditions. The same model scores differently with different prompting, more attempts, or more reasoning tokens. A benchmark number without its conditions is not a measurement; it is a claim.

Goodhart's law. A measure that becomes a target stops being a good measure. Vendors optimise for benchmarks because benchmarks sell.

It is not your job. MMLU tests multiple-choice knowledge. If your task is rewriting support emails in your house style, the correlation is anyone's guess.

How to read a benchmark claim

Ask	Because
Compared against what, when?	Vendors pick favourable and sometimes outdated rivals
Under what conditions?	Attempts, prompting and reasoning budget move scores a lot
Is the test public?	If yes, assume possible contamination
How big is the gap?	Two points is probably noise
Does it resemble my task?	Usually the honest answer is no

What to do instead

Build a small private benchmark. Twenty real examples from your actual workload, with answers you agree are right. Run candidate models against it.

It is unglamorous, it takes an afternoon, and it will tell you more than every published score combined — because it is not public, so nothing trained on it, and it is your job rather than a proxy for it.

The published numbers are useful for one thing: knowing which models are worth the afternoon.

What we addedA guide to reading a benchmark claim adversarially — what contamination, cherry-picking and undisclosed conditions do to the number in the announcement.

This article was researched and drafted with AI assistance from the sources listed below, then checked and edited by Fiqhro Dedhen before publication. How we work.