Skip to content

A benchmark score is a shortlist tool and not proof that a model fits the job.

A funnel using the example of a pump manufacturer that wants a language model to draft answers to service requests written in German. At the top are public leaderboards with many models. The shortlist leaves three candidates, which are tested on a few hundred real, anonymized service requests with recorded correct answers. Scoring covers technical accuracy, spare-part names, cost per request and latency; candidates that do worse on the company’s own cases drop out. At the bottom is the decision based on the company’s own test, not the leaderboard rank.

A benchmark works like a standardized inspection. Every model gets the same tasks, answers are scored by the same rule, and the result is a number, usually the share of tasks solved correctly. Because the conditions match, you can line up models from different providers and versions. Leaderboards collect these scores for many models.

Most benchmarks test a narrow capability. MMLU covers knowledge across 57 subject areas, GPQA hard questions in biology, physics and chemistry, SWE-bench fixing real bugs in software. Humanity’s Last Exam was built because leading models now solve older tests almost completely. Chatbot Arena measures something else: which of two answers people prefer.

For executives, a benchmark score is a shortlist tool and not proof that a model fits the job. It shows how a model does on someone else’s task set. How well it handles your own quotes, service requests or contracts only shows in a test with your own cases, called an eval.

Chapter 1 of 6

How a benchmark is built.

A benchmark consists of a fixed set of tasks, a stored correct answer for each and a rule for scoring responses. On multiple-choice questions, as in MMLU or GPQA, the chosen option counts. In SWE-bench the model has to change a real project’s code so that a described problem is fixed; automated tests check whether it worked.

Preference comparisons work without a reference answer. On Chatbot Arena, users ask their own question, get answers from two anonymous models and vote for the one they like better. The platform builds a ranking from many such pairwise votes; the 2024 paper drew on more than 240,000 of them. The platform now runs under the name Arena.

How a score was produced changes it. Whether a model had one attempt or several, which settings it ran with and which version of the test was used belong in any serious claim.

Chapter 2 of 6

Well-known benchmarks measure very different things.

MMLU (Measuring Massive Multitask Language Understanding) dates from 2020 and covers 57 subject areas, from elementary math through US history to law, as multiple-choice questions.

GPQA, from 2023, has 448 multiple-choice questions written by experts in biology, physics and chemistry. Experts who hold or are pursuing a PhD in the field scored 65 percent. Skilled non-experts reached only 34 percent despite spending more than 30 minutes on average with unrestricted web access, hence the label “Google-proof.”

SWE-bench, also from 2023, has 2,294 tasks drawn from real issues and pull requests in 12 popular Python projects on GitHub.

Humanity’s Last Exam has 2,500 questions across dozens of subjects, written by experts worldwide. Each has an unambiguous, verifiable answer that can’t be looked up quickly online. The paper appeared in Nature in January 2026.

Chapter 3 of 6

High scores can mislead.

A benchmark loses its value once the best models solve nearly every task. That saturation has happened with MMLU: according to the authors of Humanity’s Last Exam, language models now score above 90 percent on it.

The second risk is data contamination. When test questions or their answers end up in training data, the score partly measures memory instead of skill. In 2024 a team led by Hugh Zhang wrote a new math test in the style of the widely used GSM8k. On the new problems, accuracy dropped by up to 8 percent, and several model families showed signs of systematic overfitting; the leading models showed few such signs.

Third, vendors choose what they show. The 2025 study “The Leaderboard Illusion” describes how some providers could test several variants privately on Chatbot Arena before release and retract results. Publishing only the best score skews the leaderboard.

Chapter 4 of 6

A high leaderboard rank says little about your own use case.

MMLU and GPQA ask their questions in English and test university-level knowledge. Neither measures whether a model gets customer requests right when they’re written in German, say, and full of your industry’s terminology. An overall score across many tasks can also hide weaknesses in exactly the task type that matters to you.

Preference leaderboards have a limit of their own. They show which answer people would rather read. Whether it’s factually right is another matter; a fluent, confident answer can win and still be wrong.

Cost per request, latency, data protection and the provider’s reliability appear in none of these scores. For a company they often decide the outcome.

Chapter 5 of 6

The decision comes from a test built on your own operations.

Public scores help narrow many models down to two or three candidates. The decision should come from your own test, an eval. A company collects real, anonymized cases, such as customer requests or contract reviews, and records the correct answer for each. Criteria and weighting are fixed before the run. All candidates run against the same test set, and cost and latency are measured alongside quality. The test set runs again whenever the model changes.

The AI Act addresses model evaluations too. Under Article 55(1)(a) of the EU AI Act, Regulation (EU) 2024/1689, providers of general-purpose AI models with systemic risk must evaluate their models using standardized protocols and tools that reflect the state of the art. Annex XI requires them to describe their evaluation strategies, including results. That duty falls on model providers. For a company that uses such a model, it doesn’t replace a test with its own cases.

Chapter 6 of 6

Where the term stops working

A benchmark measures a model at one point in time on someone else’s task set; it proves neither fitness for a specific task nor the safety or legal compliance of a system. Evals are a company’s own tests with cases from its operations. Hardware benchmarks measure computing speed rather than model capabilities. Advertising benchmark scores without naming the test, version and date comes close to AI washing.

Frequently asked questions

What is an AI benchmark?

An AI benchmark is a standardized test that lets you compare AI models on a specific capability. Every model solves the same tasks, and the answers are scored by the same rule. The result is usually the share of tasks solved correctly.

Which AI benchmarks are there?

Well-known benchmarks for language models include MMLU for broad knowledge, GPQA for hard science questions, SWE-bench for fixing software bugs and Humanity’s Last Exam for expert-level questions. Chatbot Arena compares models based on users’ votes.

How do you compare AI models for your own company?

The most reliable way is your own test built from real cases in your operations. Public benchmarks help shortlist two or three candidates, which then run against the same test set. Quality, cost and latency are scored against criteria set in advance.

Are AI benchmarks meaningful?

For the capability they measure, yes; for use in your own company, only to a degree. Saturation, data contamination and selective publication of scores can distort results. Every score is a snapshot of one model version.

What is the difference between a benchmark and an eval?

A benchmark is a public, standardized test, while an eval is a company’s own test with cases from its operations. The benchmark makes models comparable in general. The eval shows whether a model handles the specific task well enough.

What does data contamination mean for AI benchmarks?

Data contamination means test questions or their answers have made it into a model’s training data. The score then partly measures memory instead of ability. Newly written, comparable problems expose this, because scores drop on them.

Where can you find current AI leaderboards?

Model providers publish scores with new models, individual benchmark teams post them on their websites, and platforms such as Arena, formerly Chatbot Arena, run public rankings. Every score needs a date, model version and test settings. Without them it can’t be put in context.

    ↑ ↓ select · Enter open · Esc close