Benchmark scores are the main public evidence that models are improving. They are also unusually short-lived, because the act of using a benchmark widely is what destroys it.
A benchmark is a proxy, not the goal
Every benchmark stands in for something broader that cannot be measured directly, such as reasoning or comprehension. The test is a sample, and the sample is narrow.
As long as nobody is aiming at the sample, performance on it tracks the underlying ability reasonably well. That correlation is the entire value of the measurement.
Once the sample becomes a target, the correlation weakens. Improving the score and improving the ability start to diverge, quietly at first.
Contamination is hard to rule out
Public benchmarks live on the open web, which is also where training data comes from. Test questions and their answers end up inside training corpora without anyone intending it.
Filtering helps but is imperfect, because questions get reworded, translated and discussed in forums. A model can absorb the content without ever seeing the exact string.
The result is that a strong score may reflect recall rather than capability, and distinguishing the two after the fact is difficult.
Optimising to the test is rational
Scores influence funding, attention and adoption, so teams tune towards them. This does not require dishonesty, only ordinary responsiveness to what is being measured.
Choices about prompting, formatting and answer extraction can move results noticeably without changing the model. Comparisons across published numbers are therefore less clean than they appear.
Saturation ends the useful range
When several systems all score near the maximum, the benchmark stops separating them. Remaining differences fall within the noise of the test itself.
Many older benchmarks reached this state, which is why the frontier keeps moving to harder tasks. Each replacement then begins its own decay cycle.
Saturation also hides regressions, because a test that everything passes cannot report that something got worse.
What partially replaced fixed tests
Held-out evaluations that are never published, human preference comparisons and task-specific harnesses built by the organisation deploying the model all avoid contamination by construction.
They are harder to compare across organisations, which is precisely the property that keeps them meaningful. Private measurement resists the optimisation pressure that erodes public scores.
The practical reading is that a headline benchmark number describes a model's relationship to that benchmark, and only loosely describes what it will do on unfamiliar work.