AI benchmarks are standard tests, fixed sets of questions, tasks or datasets, used to compare AI models on the same footing. They're useful for tracking progress, but scores can be inflated by saturation and data contamination, so the test that should settle a buying decision is one run on your own work.
Every AI model launch now arrives with a table of scores, and every table seems to show the new model winning. If you're choosing a tool for your business, those numbers are tempting to take at face value. This page explains what a benchmark is, how to read a leaderboard, why the scores often say less than they appear to, and what to do instead. If the underlying technology is new to you, start with our beginner's explanation of how AI works.
Benchmark, defined
Vanessa Parli of the Stanford Institute for Human-Centered AI puts it simply in an interview on benchmark saturation: a benchmark is essentially a goal for the AI system to hit, a way of defining what you want a tool to do and then working towards it. Her example is ImageNet, a dataset of over 14 million images that researchers use to test image classification. The goal is to identify as many of the images correctly as possible.
NIST's draft evaluation framework, NIST AI 200-2, gives the more formal version: benchmarking is a subset of model testing that uses standardized datasets, prompts or evaluation tasks to compare model outputs against expected answers. It says benchmarks enable consistent comparisons and have historically contributed to progress in AI development.
Both descriptions are fair. The trouble starts when a score built for comparing research systems gets read as a promise about how a tool will behave on your inbox, your product pages or your customer questions.
Reading a leaderboard without being fooled
Leaderboards rank models by their scores on one or more benchmarks. Before you trust one, check four things:
- What's being tested? A maths benchmark tells you little about writing product descriptions. Match the benchmark to the job you care about.
- How big is the gap? Stanford's 2025 AI Index found models bunching up at the top. On the Chatbot Arena Leaderboard, the Elo score difference between the top and 10th-ranked model narrowed from 11.9% to 5.4% by early 2025, and the gap between the top two shrank from 4.9% in 2023 to 0.7% in 2024.
- Is the test near its ceiling? If most models score close to the top, small differences mean very little.
- What did the score cost? The AI Index notes that OpenAI's o1 scored 74.4% on an International Mathematical Olympiad qualifying exam against GPT-4o's 9.3%, but was nearly six times more expensive and 30 times slower.
That last point gets missed all the time. A model can top a chart and still be the wrong choice if it's slower and dearer for a task a cheaper model handles well. And with the leaders so close together, the brand name on the chart matters less than how a tool fits your workflow.
Saturation: when the test runs out of headroom
Saturation means a benchmark has stopped separating good from better. In the Stanford HAI interview, Parli says most benchmarks were hitting a point where 'we cannot do much better, 80-90% accuracy'. Her example: the best image classification system on ImageNet in 2021 had an accuracy rate of 91%, and 2022 saw only a 0.1 percentage point improvement.
She also points out that some benchmarks already sit above the human baseline. On the Visual Question Answering Challenge, the top model hit 84.3% accuracy against a human baseline of about 80%. And yet, she says, a tool that meets benchmarks well can still give incorrect answers, say things you don't want it to say, and be difficult to interact with.
The research world's answer is harder tests. The 2025 AI Index reports that on benchmarks introduced in 2023, scores jumped within a year: gains of 18.8 and 48.9 percentage points on MMMU and GPQA, while on SWE-bench AI systems went from solving 4.4% of coding problems in 2023 to 71.7% in 2024. It describes traditional benchmarks such as MMLU, GSM8K and HumanEval as saturated, and lists newer tests built to stretch models: on Humanity's Last Exam the top system scores 8.80%, and on FrontierMath AI systems solve only 2% of problems.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
For a buyer, the lesson is that a headline score has a short shelf life. A test that looked hard a year ago may now say almost nothing about which tool is better.
Contamination and Goodhart's law
NIST's draft names the other big problem. It says benchmarks often measure proxies contained in datasets rather than real-world outcomes, and that models may encounter benchmark or test data during training, leading to data contamination and overly optimistic results. Because many widely used benchmarks are open and publicly accessible, NIST says they're susceptible to data leakage, task contamination, saturation and other forms of performance inflation.
It links this to Goodhart's law: measures can degrade once they become targets of optimization. Optimizing systems mainly for benchmark performance, NIST says, can produce models that perform well during testing but fail in deployment. Its conclusion is that model testing results should be complemented with testing approaches that reflect real-world conditions.
The AI Index makes a related point about what scores can hide. Even with chain-of-thought reasoning, it says, systems still can't reliably solve problems such as arithmetic and planning that have provably correct solutions, especially on instances larger than those they were trained on.
Run your own test instead
NIST suggests one way to reduce leakage and contamination: build in-house or custom benchmarks for targeted applications. That's the move I'd make for any business choosing an AI tool. You don't need a lab. You need a small, honest test.
- Collect real examples of the work you want the tool to do: emails, briefs, support questions, product pages. NIST's draft stresses that test data should reflect the intended use where feasible.
- Write down what a good answer looks like before you run anything. NIST recommends clearly defining what's being measured and how results will be interpreted.
- Run the same inputs through each tool you're considering, more than once. NIST suggests multiple trials to account for variability in model outputs.
- Compare against a baseline, such as your current process or a person doing the task. NIST's draft calls for appropriate reference points for comparison.
- Score what matters to you: accuracy, tone, speed and cost, not only the headline metric.
If tone matters, put your own voice into the test; our guide on training AI to write in your brand voice shows how to set that up. And if you're comparing tools that handle images, audio and text together, it helps to understand how multimodal AI differs first, since a text-only benchmark won't tell you much about them.
Treat benchmarks as a decent first filter and a poor final answer. Use the leaderboard to build a shortlist, then let your own work pick the winner.
