Top agents by capability

A table ranking models by their score on one benchmark. The same fixed task set is run against each model and the results are sorted. Because each leaderboard covers a single capability, a model that tops coding can sit mid-table on maths.

It depends on the coding work. For fixing real issues in an existing codebase, look at SWE-bench Verified; for multi-language editing, the Aider polyglot leaderboard. The leading models usually sit within a couple of points of each other, so price and latency often decide it. (Aider)

A set of multiple-choice science questions written by PhD holders in biology, physics and chemistry. They are hard enough that skilled non-experts with unrestricted web access score around 34%. It is used to test graduate-level reasoning rather than recall. (GPQA paper)

Because the benchmarks are close to saturation. Once several models sit within a point or two, the gap is inside the noise of a single evaluation run and the ranking stops picking a winner. Selection then moves to price, latency, context length and tool-calling reliability.

Some come from the labs themselves, which is why independent runs matter. Epoch AI publishes its own evaluations under CC-BY, and the Aider polyglot leaderboard is published under Apache 2.0. New models are scored as they ship, so any ranking is a snapshot. (Epoch AI)

Mostly by accident. Benchmark questions leak into training data over time, which inflates scores without improving the model. Labs also choose which benchmarks to report and which prompting setup to use. Independent runs and held-out test sets are the usual defence.