Can You Trust Benchmark Scores?
How to read an AI score chart
① A benchmark score is only "how many questions on a fixed test did it get right," not its performance on your work. Test questions themselves can contain errors, and the same score doesn't mean the same thing under different conditions.
② There are routes by which scores get inflated: test questions leaking into training data, trying many variants privately before release and keeping the best, and releasing a model tuned specifically for a ranking. There are real cases.
③ Read a score chart by first asking who measured it, under what setup, and which version — then make the final call by comparing directly on a few tasks like your own.
① Getting started — the score chart that comes with every new model
Every time a new AI model is announced, a bar chart follows: "No. 1 on the math benchmark," "record coding score." As an investor, you'll often see such scores turn into stock news too. Yet what these numbers measure, and how far to trust them, is rarely explained. This piece lays out how to read them.
② What does a benchmark measure?
A benchmark is a test made of a fixed set of questions and answers. Models solve the same questions and the share answered correctly becomes the score. The strengths are clear: every model can be compared on the same questions, and you can track change over time.
The limits are clear too. A high score doesn't mean it does real work well. A model that answers multiple-choice knowledge questions accurately may differ when organizing a long report or handling your company's documents. The starting point is that a benchmark is "a grade on this test," not "performance on your job."
③ Five reasons scores mislead
| Reason | What happens |
|---|---|
| ① Errors in the questions | The test questions or answer key are wrong. A study where people re-reviewed the widely used MMLU found errors in 6.49% of questions, and in the virology subject, 57% of the questions analyzed (Gema et al., 2024). With errors, model rankings can flip |
| ② Saturation | Top models cluster near a perfect score and discriminating power disappears. The gap between 98 and 99 can be closer to errors in the questions or chance than to model skill |
| ③ Contamination | If test questions or answers leak into training data, the model gets them right by memorizing, not by solving. Since training gathers text from across the internet, it's hard to verify from outside how much leaked |
| ④ Differences in setup | The same score can mean different things. The number moves a lot depending on how many examples were shown, whether tools (search, a calculator) were used, and whether the best of several attempts was picked |
| ⑤ Tuning to the score | When the score becomes the goal, people optimize for raising the score. That can mean releasing a model specially tuned for a ranking, or privately testing many variants and publishing only the good ones |
④ Two real cases
1. A model tuned for the ranking — Llama 4 (April 2025)
When Meta announced Llama 4 Maverick, it said an "experimental chat version" scored 1417 on LMArena (a ranking site where people vote between two models' answers). But the model on the ranking was a version tuned for human preference, different from the model released to the public. LMArena stated that Meta should have made that clearer. According to reporting, the experimental version gave longer answers with more emojis than the public one, and when the unmodified public version was evaluated separately, it came in 32nd (The Register, Slashdot). Meta explained that it experiments with all types of custom variants. The score was real, but it wasn't the score of the model users would actually use.
2. Testing many variants privately — the "Leaderboard Illusion" study
In a paper analyzing the same ranking site (Chatbot Arena), a group of researchers pointed out that some providers can privately test many variants before release and put out only the good one. They reported that Meta tested 27 private variants in the lead-up to the Llama 4 release. In an experiment submitting two identical checkpoints of the same model, they also observed a 17-point score difference. The implication is that part of a ranking can reflect number of attempts and chance rather than skill (Singh et al., The Leaderboard Illusion, 2025).
These two cases don't mean "every benchmark is rigged." Most benchmarks are useful. They show that a high score alone can't establish "the model I'll use is better." The figures reflect the time of the study and reporting, and the ranking site's policies may have changed since.
⑤ Six questions for reading a score chart
- Who measured it? The model maker's own announcement, or an independent third-party evaluation? Self-reported results make it easy to choose favorable conditions.
- Under what setup? Are the number of examples, tool use and number of attempts stated? If not, you can't compare.
- Is it a like-for-like comparison? Check whether all the models in the chart were run under the same setup, or only the competitors under a different one.
- Which version's score is it? Confirm whether it's the public version you actually use, or an experimental or internal one.
- Does this test look like my work? A math score tells you nothing about report-summary quality. Look at a benchmark of the type closest to your task.
- Is it saturated? If the top models all sit near a perfect score, that benchmark can't show differences.
Pick about 5 jobs you'd really hand over and give them to every candidate model under the same conditions. Five questions is statistically crude, but the difference on your own use case is far more direct than a benchmark bar chart. When scores are close, this test often decides it. For how to run it, see the "30-minute self-test" in comparing AI tools, and for which model to pick for which purpose, "which model, when" in what's different about reasoning models.
⑥ Investor angle — a score is an event; earnings are the evidence
A new model's score is used in markets as an event. Related stocks sometimes move right after an announcement. But "No. 1 on the score" isn't revenue or profit. Whether a company's AI spending pays back shows up not in a score chart but in the numbers and guidance in its earnings. How to read AI demand in big-tech earnings is covered in the big-tech earnings interpretation guide, and a case where the headline number is good but the number beneath it (margin) is the new information is covered in Nvidia's Q2 FY2027 earnings, dissected. Both are ways of looking at how a number was produced, not just at the one big number announced.
⑦ Practical checklist
□ Was the score produced by the maker or by an independent third party?
□ Are the setup details — number of attempts, number of examples, tool use — disclosed?
□ Were all compared models measured under the same conditions?
□ Is it the score of the public version I'll actually use?
□ Does this benchmark resemble the work I'll hand over?
□ Aren't the top models all bunched near a perfect score?
□ Before the final call, did I compare directly on a few of my own tasks?
※ The cases and figures here are based on research and press reports published in 2024–2025 and may have changed since. This is not an assessment of any particular company or model, and investing-related content is not investment advice. Investment decisions and responsibility rest with the investor.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.