- When a lab says its model beats the others, the benchmark behind that claim is often contaminated, saturated, or gamed, so the number means less than it sounds.
- Once a benchmark becomes a marketing target, labs optimize for the score rather than the underlying ability. That is Goodhart's Law in action.
- Serious evaluators now ignore the saturated classics and watch contamination-resistant tests like Humanity's Last Exam and FrontierMath, where even top models still fail most of the time.
The Setup
Every few weeks a lab announces a model that beats GPT-4 or tops some leaderboard, and the headline gets treated as fact. Most of those numbers are close to meaningless. The benchmarks behind them are saturated, leaked into training data, or quietly optimized for. Knowing why is the difference between reading AI news and understanding it.Problem One: Saturation
The classic benchmark, MMLU, a broad multiple-choice test, is finished as a frontier tool. In 2026 every leading model scores above 88 percent on it, and the gaps between them sit inside measurement noise. When five models all score 89 to 92, ranking them by that number is like ranking sprinters with a stopwatch that only reads whole seconds. The test stopped discriminating, but the marketing did not stop using it.Problem Two: Contamination
Here is the one most people miss. Benchmarks are public, and models train on the public internet, so the test questions often end up in the training data. The model then scores by remembering, not reasoning. Studies have found contamination as high as 45 percent on common benchmarks, and when researchers stripped leaked examples out of a math test, accuracy dropped about 13 percent, a floor and not a ceiling. A contaminated score measures memory, and memory is not intelligence.Problem Three: Goodhart's Law
There is a rule that governs all of this: when a measure becomes a target, it stops being a good measure. Once benchmark scores became the headline metric labs market on, the incentive shifted from building general capability to optimizing for the specific test. Teams pick checkpoints that score well on the benchmark rather than ones that are genuinely better. The number goes up, the marketing improves, and the real-world ability barely moves.What Experts Actually Watch
Serious evaluators have moved on to contamination-resistant tests. Humanity's Last Exam, a 3,000-question gauntlet built to be Google-proof, is now the frontier yardstick, and the top model scores only around 37 percent, with most below 30. FrontierMath uses brand-new, unpublished problems written by more than 70 mathematicians under strict secrecy. LiveCodeBench filters to coding problems created after a model's training cutoff. The common thread is simple: a test the model could not have seen before.What It Means For Investors
This is where it pays off. A new state of the art press release is not a reason to reprice a stock or rip out your stack. Benchmark wins are now a marketing surface, and the gap between leaderboard and real-world usefulness has widened. The signal that matters is performance on held-out, contamination-resistant tests, and even more, whether the model actually works on your task. Treat round-number benchmark claims as advertising until proven otherwise.FAQ
So are all benchmarks useless?No, they are useful when they are fresh and uncontaminated. The problem is the famous, saturated ones that dominate marketing. Held-out tests like FrontierMath or LiveCodeBench still carry real signal, at least until they leak too.
How can a model cheat if no one fed it the answers on purpose?
It usually isn't deliberate. The test questions are public, so they get swept into the training data along with everything else online. The model memorizes them by accident, and the benchmark can no longer tell memory from reasoning.
What should I do with an AI benchmark claim?
Discount it. Ask which benchmark, whether it is contamination-resistant, and how the model does on tasks it could not have memorized. And test it on your own use case, that is the only benchmark that pays your bills.