Compare the task you care about, with the source and evaluation setup visible.
A separate ranking for each benchmark
Keep native units, versions and task settings. An arena rating, pass@1 and resolved-task percentage have different meanings.
Agentic, Design and Coding stay distinct
Agentic uses LiveBench 2.0 Long Agentic tasks. Coding and Design use their own suites in the same release. Model settings are retained as published.
Evidence belongs beside the score
Show source-reported sample counts and uncertainty when available. Missing values stay empty; untested models stay unranked.
Match image and video settings
Separate generation from editing, and text-to-video from image-to-video. Arena overall cohorts can include different model defaults. Check the source before comparing audio, resolution, length or quality tiers.
Changes need a stable reference
Only compare trend points within the same benchmark version and setup. A new evaluation version starts a new series.