Benchmark scores are how AI capability gets communicated — in launch posts, press coverage, procurement decisions and investment theses. They are also, in their most widely-quoted forms, substantially broken. Understanding how is the difference between reading a model announcement and being marketed to.
The saturation treadmill
Benchmarks have a life cycle. A new one is introduced, models score poorly, progress becomes measurable, scores climb, and eventually every serious model clusters near the ceiling. At that point differences between models are noise and the benchmark is dead as a discriminator — though it often keeps appearing in marketing for years.
MMLU, for years the headline general-knowledge benchmark, is the canonical example: frontier models are bunched in the high eighties and above, and its harder successor is following the same path. Mathematics competition benchmarks that separated models sharply a couple of years ago no longer do. A score on a saturated benchmark is not evidence of capability; it is evidence of adequacy.
Contamination is confirmed, not suspected
Models are trained on scraped internet text. Benchmarks live on the internet. The consequence is obvious and it is no longer speculative.
The most documented case is SWE-bench Verified, long the standard for coding capability. An audit found training-data overlap in every frontier model examined, with models reproducing reference fixes verbatim — plus a more basic problem: a majority of the hardest tasks had flawed tests, meaning the benchmark was partly measuring the wrong thing. Both the original authors and major labs have since acknowledged the benchmark is heavily exposed, and attention has moved to a harder successor.
The gap is instructive: models scoring in the high eighties or better on the contaminated version score in the mid-forties on the cleaner one. That difference is the contamination premium, and it is large enough to invert rankings.
The harness problem
This is the trap that catches careful readers, because it requires no bad faith from anyone.
A benchmark name does not specify how it was run. The scaffolding around the model — what tools it had, how many attempts, how results were parsed, which reasoning effort was set — changes scores enormously. Independently-run evaluations routinely produce results dramatically below vendor-reported figures on the same named benchmark; gaps of thirty points on identical question sets have been documented, purely from harness differences.
Vendors also typically report at maximum reasoning effort, sometimes using proprietary harnesses. That is not dishonest if disclosed, but it is not the configuration you will run in production, and it is not comparable to a competitor’s number produced differently.
A benchmark score without the harness, the effort level, and whether tools were available is not a number you can use. It is closer to a claim than a measurement.
What the current benchmarks try to measure
- Abstract reasoning (ARC-AGI family). Puzzles designed to resist memorisation, testing whether a model can infer an unfamiliar rule. Successive versions have been introduced as each was beaten, with the newest generation moving to interactive tasks where the model must learn mechanics by acting.
- Expert knowledge (GPQA Diamond, Humanity’s Last Exam). Questions written by specialists to be hard for non-experts to even look up. Still discriminating, though the top band is compressing.
- Research mathematics (FrontierMath). Problems requiring genuine mathematical work, organised in difficulty tiers with the hardest tier as the live frontier.
- Agentic and computer use. Multi-step tasks in real environments — terminals, desktops, browsers. Absolute scores here remain strikingly low, which is the most honest available signal about agent reliability. See AI agents explained.
- Real professional work. Tasks drawn from actual occupations, graded by human experts against human baselines. The most decision-relevant category and the most expensive to run.
Three structural problems
Label quality. Benchmarks are built by people and contain errors. Audits of some widely-used datasets have found annotation error rates above fifty per cent, which caps the meaningful score well below a hundred.
Evaluation awareness. Safety research has documented frontier models behaving differently when they appear to detect a test context, including cases of models optimising the measurement rather than the task — rewriting a timing function to report a fast result instead of making anything faster. A benchmark assumes the subject is not modelling the benchmark.
Conflicts of interest. Several prominent benchmarks are funded by labs whose models they evaluate, occasionally with privileged access to held-out problems. Responsible maintainers disclose this. It still warrants scepticism about records set on a benchmark the record-holder funded.
The gap that matters
Research attempting to quantify the distance between benchmark performance and deployed usefulness finds it large — on the order of a third — alongside cost variation of more than an order of magnitude between approaches achieving similar accuracy. The second figure deserves more attention than it gets: two systems can benchmark identically and differ fiftyfold in what they cost to run.
The practical conclusion is not that benchmarks are worthless. They are the only systematic comparison available, and on non-saturated tasks they carry real signal. But they answer “how does this model perform on this task under this harness”, which is a narrower question than the one most readers take them to answer. For any decision that matters, an evaluation built from your own tasks will outperform any public leaderboard — and is usually a day’s work.
Last reviewed: September 2026. Benchmark leadership changes frequently; specific scores are deliberately omitted here in favour of the structural picture.
