Read a benchmark as one column in a table, not a verdict. A model that leads on a reasoning benchmark may be slower or more expensive for the short classification calls that make up most of a production workload. Two models can be separated by a point or two while differing by a factor of ten in price, and the leaderboard will not tell you that.
The directory shows benchmark values where a public source or model card provides them, and names the source. Cells are highlighted to show the best number in that row only. A highlighted cell is not a recommendation.
The reliable test is your own prompt, with your own data, graded by someone who knows what a good answer looks like. Use benchmarks to build a shortlist; use your prompt to pick from it. See deciding after a model comparison.