Legacy benchmarks are saturated, and the people who design them have had to reinvent the job: questions pre-screened by the models themselves, cash bounties for authors, private holdout sets. But the craft has discovered the harder problem isn’t question difficulty — it’s who grades, who funds, and who decides which scores make the leaderboard.
To get into the dataset, a question first has to fail. That was the governing rule behind Humanity’s Last Exam, the 2,500-question benchmark released in January 2025 by the Center for AI Safety together with Scale AI. Anyone submitting a problem had to run it past a battery of language models first; if the models answered correctly, the question was thrown out. Only the ones the systems got wrong moved on to two rounds of human review. Authors were paid: a total prize pool of $500,000, $5,000 each for the first fifty accepted questions, $500 for the next five hundred.
It is a procedural detail that says more about the state of AI evaluation than most commentary does. The thing being measured here is large language models — systems trained on enormous text corpora that produce natural-language answers — and the problem is that there is no longer an obvious way to measure them. Benchmark designers no longer write “a hard test.” They write a test against an adversary that already knows most of what has ever been written down. The model isn’t just the object of measurement; it’s also the first filter deciding what goes into the ruler.
Why the job changed
Stanford HAI’s AI Index 2025 makes the urgency numerical. On what were, until recently, the frontier benchmarks — MMMU for multimodal reasoning, GPQA for PhD-level questions, SWE-bench for real software engineering tasks — models improved between 2023 and 2024 by margins that don’t look like incremental progress: +18.8 percentage points on MMMU, +48.9 on GPQA, +67.3 on SWE-bench. The report notes this “pushes researchers to propose harder benchmarks,” and explicitly cites Humanity’s Last Exam as a response to the saturation of popular tests.
Saturation means something specific. Once the best models clear 90% on a test, the gap between two systems lives in the noise. Wrong answer keys, ambiguous phrasing, quirks of automated grading start to weigh as much as actual capability. The benchmark stops ranking and starts drawing lots.
The natural response — write harder questions — is the one the community took. Humanity’s Last Exam is 41% mathematics, 11% biology and medicine, 10% computer science and AI, 9% physics; roughly a quarter multiple choice, roughly 14% with multimodal components. The authors held back a private, unpublished set of questions specifically to measure how far models are overfitting to the test rather than genuinely improving. Ten years ago that precaution wouldn’t have been worth writing into a paper. Today it’s standard practice.
How long they actually last
Here it pays to be precise, because the standard line — benchmarks burn out in months — is only sometimes true. At launch, GPT-4o scored 2.7% on Humanity’s Last Exam and o3-mini in its high configuration reached 13.4%. By October 2026, with nearly three hundred models evaluated, the top of the board sits around 59%. That’s almost two years, and the test is not saturated. The paper reached Nature in late January 2026.
FrontierMath, the mathematics benchmark built by Epoch AI from original problems never published anywhere, tells a different story: by mid-2026 its Tier 1-3 levels were being cleared at roughly 87%. And this is the uncomfortable observation for anyone assuming the culprit was training-data contamination — the phenomenon where test questions end up, verbatim or paraphrased, inside the corpus a model was trained on. FrontierMath was engineered precisely to rule that out: unpublished problems, never online. It got saturated anyway.
Contamination remains a documented systemic risk — a recent systematic review catalogues 55 studies and proposes a four-level taxonomy: exact, syntactic, semantic, task-level; Balloccu and colleagues in 2024 produced the first systematic analysis of closed-source models, covering 255 papers that used GPT-3.5 and GPT-4. But no source quantifies how much of the saturation is contamination and how much is real progress. Anyone claiming otherwise is guessing.
The dirty work: auditing the auditors
The least-told part of this job is that a benchmark, once published, becomes an audit target itself. In 2025 FutureHouse examined the text-only chemistry and biology questions in Humanity’s Last Exam and estimated that around 30% of the reference answers were wrong. The benchmark team partially replicated the finding and went on to launch HLE-Rolling, a continuously updated version. Worth pausing on: the dataset used to decide whether a model is any good contained, in one of its sections, a significant share of incorrect answers. And an outside group is what caught it.
Nor is there, today, a single HLE leaderboard. Different evaluation harnesses and different subsets — full set, text-only, a “clean” thousand-question subset — produce numbers that are not comparable with each other. Stitching percentages from different boards into a single progress curve is a common error and it does not hold up.
The structural conflict
Then there’s the question of who pays. Epoch AI stated in a conflict-of-interest disclosure that OpenAI commissioned 300 FrontierMath problems, owns them, and has access to the statements and solutions — with the exception of a declared 50-problem holdout. The disclosure came late: December 2024, and only in the final version of the paper. Some of the mathematicians who had contributed said so publicly. This is not an accusation of cheating. It’s a description of an arrangement where the party being evaluated funds the party doing the evaluating, and where information about that arrangement arrived after the headlines did.
The document that made the most noise in this direction is The Leaderboard Illusion, 68 pages posted to arXiv on 30 April 2025 by a group from Cohere Labs, Princeton, Stanford, Waterloo, MIT, AI2 and the University of Washington — authors include Shivalika Singh and Sara Hooker — later accepted to the NeurIPS 2025 Datasets & Benchmarks track. It documents undisclosed private testing in human-vote comparison arenas: 27 private variants tested by Meta ahead of the Llama 4 launch, relative gains of up to 112% obtainable through access to arena data, and 205 out of 243 models silently deprecated. The authors describe starting the study after submitting their own open-weight models in October 2024 and contacting the organisers the following month.
In the Llama 4 case, in April 2025, the model topping the arena was not the one subsequently released to the public. LMArena issued an apology, while disputing what it called factual errors in the paper.
Who quits
On 13 March 2025, Hugging Face retired the Open LLM Leaderboard. In two years it had evaluated more than thirteen thousand models. The announcement came from Clémentine Fourrier; the space was archived and no further updates are planned.
Thirteen thousand models in twenty-four months is the figure that best explains why this work became unsustainable in the shape it had. Not because the questions got hard — because of volume. A free public leaderboard that evaluates every system that shows up is a piece of infrastructure somebody has to maintain, with compute costs that scale with the number of entrants and no revenue that scales with them. Designing benchmarks, right now, looks less like educational测 measurement theory and more like maintaining a public road while the trucks using it keep getting heavier.