01What we cite
Public benchmarks and leaderboards (for example SWE-bench, LMArena, Artificial Analysis and MLPerf), peer-reviewed and preprint research, model system cards, and results published by the companies that make the products.
A company’s own numbers are labelled as self-reported. Where an independent evaluation exists, we cite it alongside.
02How we credit results
Every Benchmarks story names the benchmark, who ran it, the date of the results and a link to the original. Charts carry the same credit. We never present someone else’s numbers as our own testing.
03What we check
Before publishing we read the original source, confirm the figures we quote, and note which model or product version and which setup was evaluated. If the setup differs between results, we say so instead of ranking them side by side.
04What a benchmark can’t tell you
A benchmark measures one set of tasks under one setup. We explain what was measured and what wasn’t, and we don’t treat a single score as proof that one product is better for everyone.
05Updates and corrections
Leaderboards change. When results we reported are updated or withdrawn, we update the story, log the change and, if we got something wrong, publish a correction.