How to read AI benchmarks without being fooled

In February 2026, OpenAI quietly retired the most-quoted test in AI coding. For a year and a half, nearly every model launch bragged about its score on SWE-bench Verified, a benchmark built from real bugs in real open-source projects. OpenAI helped create it. Then its own audit found that most of the hardest problems had broken grading, and that frontier models appeared to have seen some of the answers during training. The scoreboard everyone pointed at was measuring, in part, memory and luck.

This matters beyond coding. Every time you read that a new model “beats GPT-5 on benchmarks” or “scores 92% on a PhD test,” you are being handed a number with a lot of fine print nobody reads aloud. This post is that fine print, in plain language: what the big tests actually measure, the three ways their numbers go wrong, and a simple way to judge any new model yourself in about ten minutes.

Every AI headline leans on a handful of tests

A benchmark is just a fixed exam. Someone writes a big set of questions with known answers, every model takes the same exam, and the percentage correct becomes the score you see in charts. The idea is fair on its face. The same questions for everyone means the numbers should be comparable.

Four exams do most of the work in launch announcements. MMLU is a giant multiple-choice quiz across 57 subjects, from law to elementary math. GPQA is 448 science questions written by PhD holders and designed so that searching the web barely helps. SWE-bench Verified gives the model 500 real bug reports and asks it to fix the actual code. And LMArena is not an exam at all: it shows volunteers two anonymous answers to the same question and lets them vote for the one they prefer.

Each one measures something real. The trouble starts when a number from one of these exams gets treated as a general grade for intelligence. Three specific things break, and they are worth knowing by name, because they explain almost every case of a model that looked great on paper and disappointed in person.

Models can memorize the answers

Models learn from enormous scrapes of the internet, and the internet contains the exams. The questions, the answers, and thousands of pages discussing both. When test material leaks into training data, researchers call it contamination. The model is not reasoning its way to the answer; it has effectively seen the answer sheet.

The cleanest demonstration came from a team at Scale AI. They rebuilt a popular grade-school math test from scratch: 1,000 brand-new problems, written by humans, matched in style and difficulty to the famous GSM8k benchmark, and guaranteed to be in nobody’s training data. Then they ran the leading models on both versions. On the new test, some model families scored up to 8 percentage points lower than on the original, with whole families overfitting across almost every size. The models had partly memorized the public test. The math ability was real; the score just overstated it.

The same problem helped kill SWE-bench Verified. In its February 2026 audit, summarized well in this analysis, OpenAI found models solving tasks rated nearly impossible, and details from the projects’ release notes surfacing in the models’ reasoning even though the task never mentioned them. The benchmark tasks come from public code repositories, and public is exactly what training scrapes eat.

The takeaway is not that every score is fake. Frontier models showed little overfitting in the math study. It is that a score on an old, public test is a ceiling, not a promise. The real ability is somewhere at or below the number.

The tests keep getting too easy

A test only tells models apart while some of them fail it. MMLU was genuinely hard when it was published in 2020: the best model of the day beat random guessing by about 20 percentage points, and the authors wrote that the best models still needed substantial improvement in every one of the 57 subjects. Today, top models cluster above 90%, which is a big part of why researchers built Humanity’s Last Exam, a deliberately brutal 2,500-question replacement where frontier models still score low. Benchmarks have a life cycle: hard, then useful, then saturated, then retired.

GPQA tells the same story faster. When it launched in late 2023, PhD experts scored about 65% in their own fields while GPT-4 managed 39%. Within two years, frontier models sailed past the expert baseline; Epoch AI’s independent testing clocked one at 87% on the benchmark’s hardest slice. The question drifted from “can AI do science?” to a routine line in a launch table. And SWE-bench Verified spent its last six months pinned between 74.9% and 80.9% at the top, a plateau so flat that nobody could say whether the remaining failures were the models’ fault or the test’s.

Saturation has a sneaky consequence for readers: near the ceiling, small differences stop meaning anything. When three models score 89%, 90%, and 91% on a quiz with a few thousand questions, the ranking can come down to a handful of items, some of which are broken. Which brings us to the third problem.

Some of the questions are simply wrong

Benchmarks are built by people, usually quickly, often by pulling questions from the web. Mistakes get in. A team of researchers manually re-checked a slice of MMLU and estimated that 6.49% of its questions contain errors: wrong official answers, multiple correct options, or questions that make no sense. In the worst subject, virology, they found problems in 57% of the questions they analysed. On those items, a model can be marked wrong for being right.

Coding benchmarks have it worse, because grading is automatic. A fix counts as correct only if it passes the project’s tests, and tests can be unfair. SWE-bench Verified existed precisely because the original SWE-bench was noisy; OpenAI paid human reviewers to filter it down to 500 vetted tasks. Even so, the 2026 audit of its 138 hardest remaining tasks found 59.4% had flawed tests or unclear problem statements, including tests that rejected perfectly good fixes for not matching one specific implementation.

Hold those two numbers side by side and the lesson writes itself. If several percent of a test is broken, then a two-point gap between two models is inside the noise. Treat close scores as a tie.

The popularity contest can be gamed

Because fixed exams age badly, many people prefer LMArena, the blind taste test. You ask a question, two unnamed models answer, you vote for the answer you like, and only then are the names revealed. The votes pile up into a public ranking. Fresh questions arrive daily, so there is nothing to memorize. It is a genuinely clever design, and it still has a soft spot: it measures what people like, and what people like can be manufactured.

The clearest case arrived in April 2025. Meta submitted a special version of its Llama 4 Maverick model to the arena, tuned to please voters, and it climbed to #2 in the world. The version actually released to the public behaved differently, and when LMArena ranked it under the same rules a few days later, it landed at #32. Same name, thirty places apart. The arena changed its rules in response.

A later study, aptly titled The Leaderboard Illusion, documented the structural version of the same advantage. Big labs privately tested many variants and kept only the best result; Meta alone tried 27 private variants before the Llama 4 release. The study also estimated Google and OpenAI models together received about 40% of all arena votes, while 83 open-weight models shared under 30%, and that this extra exposure alone produces large ranking gains. None of this means the top of the leaderboard is fraudulent. It means a rank is partly a measure of how hard someone worked the leaderboard.

What a score can still tell you

After all that, it would be easy to conclude that benchmarks are worthless. They are not. You just have to read them the way you read a used-car listing: true facts, favorably arranged. Four habits recover most of the signal.

Trust big gaps, ignore small ones. A model that scores 30 points higher than another on a hard test is very likely better at that kind of work. A model that scores 2 points higher is tied. Between broken questions, memorized answers, and ceiling effects, single-digit gaps rarely survive contact with reality.

Prefer tests the model could not have studied for. Newer benchmarks, private question sets, and freshly written tasks beat anything that has been on the internet for years. When a launch quotes only elderly benchmarks, that is itself information.

Check who ran the test. A vendor grading its own model picks the settings, the attempts, and the framing. Independent reruns routinely land a few points lower. Not scandalous, just human. Read vendor numbers as a best case.

Match the test to your task. A coding benchmark says nothing about writing quality. An arena rank says little about factual accuracy. If you mostly summarize documents, a model’s math-olympiad score is trivia. We keep a running, plain-English account of which new models actually matter in our monthly roundup, and the same skepticism applies there: benchmarks flatter.

Ten minutes with your own tasks beats any leaderboard

Here is the practical part. You do not need a benchmark to pick a model. You need five of your own tasks and ten minutes, because the only exam that matters is the one your week keeps setting.

This is the household version of what serious teams do formally. If you run AI in a product, the same idea grows into a real test suite with saved cases and pass criteria; we wrote a hands-on guide to that in how to write your own AI eval. And if you just want a starting shortlist before you test anything, our task-by-task comparison in ChatGPT vs Claude vs Gemini was built exactly this way: same prompts, side by side, judged on the output.

The next time a headline says a model “crushed the benchmarks,” you now know the three questions to ask. Could it have memorized the test? Is the test too easy to tell models apart? And who checked the answer key? Most launch-day numbers wilt under at least one of them. Your own five tasks never do.