AI benchmarks are breaking down faster than anyone can replace them

Mark Chen describes the situation plainly: the field is in an evals crisis. The canonical benchmarks, the ones researchers have used for years, are fully saturated. And the problem compounds itself: once an evaluation is released publicly, Chen says, it is already not a good one. Contamination and overfitting begin immediately.

Elon Musk puts a sharper edge on the same diagnosis. Current evaluations are so massively overfit, he argues, that the scores cease to reflect true capability. The models overfit to the benchmarks, and the numbers stop meaning what they claim to mean. That observation is not new, but the scale at which it now operates is. Benchmark saturation used to be a lagging problem. It is now nearly simultaneous with model release.

The speed of the underlying progress is what makes the measurement gap so hard to close. On the Apex benchmark, Brendan Foody notes, the frontier model score jumped from one percent to 40 percent in 12 months. Swyx, following a similar trajectory, estimates that the Frontier Code 2026 benchmark will hit roughly 80 percent by year’s end. Sergiy Nesterenko, describing a retail AI benchmark his team built, reports that it remains far from saturated, with a rough estimate putting human-level performance at about 10 times current model scores. These are not the same benchmarks, but the pattern is consistent: the gap between what models score today and what the instrument was designed to measure is collapsing faster than new instruments can be constructed.

If you want to evaluate the capabilities of a model, what it can do after running for a month, the only way to be fully sure is to actually run it for a month. Noam Brown

Nathan Labenz, citing work from METR, puts a number on the underlying rate. The doubling time for AI task length is running at a little under four months, which implies an eight- to twelve-fold increase over the course of a single year. METR has itself noted publicly that it is struggling to construct tasks long enough to evaluate the models at all. Research published by METR confirms the difficulty: existing benchmarks saturate quickly and may be too artificial to indicate real-world capability reliably.

Noam Brown identifies why the task-length problem is not merely technical. Modern models, he explains, when scaffolded reasonably well, can think for weeks before performance plateaus. That creates an evaluation problem with no clean solution. “If you want to evaluate the capabilities of a model, what it can do after running for a month, the only way to be fully sure is to actually run it for a month.” Evaluation at that scale is not something a benchmark leaderboard can absorb. Brown also notes that benchmark scores can be inflated by scaffolding multiple model runs together, which makes cross-model comparisons unreliable even when the benchmark itself has not yet saturated.

The design flaws run deeper than speed alone. Richard Socher points to a structural bias built into most benchmarks: they are implicitly designed to cap intelligence just above human level. The result is that performance appears to plateau after crossing the human threshold, but the plateau is an artifact of how the instrument was built, not evidence of a genuine capability ceiling. Socher also flags a more basic bug in multiple-choice instruments like the Graduate-Level Google-Proof Question Answering benchmark, known as GPQA: changing the order of answer options should not affect model scores, but it does. An evaluation sensitive to answer ordering is not measuring what it purports to measure.

The implications extend beyond research labs. Brown observes that current safety frameworks, including preparedness frameworks and responsible scaling policies, do not account for test-time compute. They ask what a model can do, but model capability is now a function of how much compute is put into inference. A framework that ignores that variable is evaluating a moving target with a fixed ruler. Martin Casado notes that a given model stays relevant for roughly three to nine months before being superseded, which means the window for any evaluation to remain meaningful is narrowing even as the cost of building a valid one rises. The field is not short of ambition for better benchmarks. It is short of time to build them.