Methodology
Each job is ranked on one public board. Other boards are support, not a second vote that can overturn the leader.
The boards
- Coding & Software: LMArena Code. Also checked: LMArena Code, SWE-bench Pro, LiveCodeBench, Artificial Analysis Coding Index.
- Legal & Long Documents: Vals LegalBench. Also checked: Vals LegalBench, LegalBench, BigLaw Bench.
- Writing & Marketing: LMArena Creative Writing. Also checked: LMArena Creative Writing, LMArena Text overall, EQ-Bench Creative Writing.
- Reasoning, Math & Science: Humanity's Last Exam. Also checked: Humanity's Last Exam, GPQA Diamond, FrontierMath, Artificial Analysis Intelligence Index.
- Speed & High Volume: Artificial Analysis output speed. Also checked: Artificial Analysis output speed, Artificial Analysis latency, Artificial Analysis price.
- General & Everyday: LMArena Text overall. Also checked: LMArena Text overall, Artificial Analysis Intelligence Index, LMArena Text hard prompts.
Two checks
A daily job re-reads the boards. A new number one is noted the first day and replaces the leader only if the next run names it again. A one-day blip does not change the pick.
How sure the label is
Close race means the runner-up's score is within 1% of the leader on the primary board. Otherwise, when more than one listed board was checked, the label is how many of them agree. When only one was checked, the label is Leads followed by that board's name.
What gets published
A fact-check runs before any ranking is published. A number in a reason has to appear in the stored evidence or the price. The runner-up cannot outscore the leader on the primary board. Privacy and price notes older than 45 days are held, and the card says it is being re-checked instead of showing the stale line. A change in how a category is ranked is logged on Changes as a methodology update and is not sent as an alert.
Benchmarks measure a public task, not your files. Treat the top two picks as a shortlist and try both on a sample of your own work.