Kaggle launches Arenas, where your votes rank AI models head to head

Kaggle launched Arenas, head-to-head matchups where AI models take on the same challenge and you decide which one comes out on top. Kaggle's pitch is that benchmarks tell you how models score, and Arenas let you judge the results yourself.

Each matchup shows two anonymous outputs for the same prompt, side by side, whether that's a video, a web page, an image or text. You vote A is better, B is better, Both are good or Both are bad, and every vote feeds the leaderboard.

The Arenas page lists 15 arenas so far. HeyGen's Code2Video Arena tests models on prompt adherence when generating HyperFrames code, and MineBench Arena covers 3D Minecraft voxel builds across 27 prompts. Kaggle's own include Pelican Arena, where models draw an SVG of a pelican riding a bicycle, Chart Arena, which turns a small dataset and a brief into one self-contained SVG chart, Floorplan Arena, where models furnish hand-drawn room sketches, and Landing Page Arena, a one-page website for a small business. There are text arenas too, for jokes, ELI5 answers and novel opening lines.

Several arenas set a pass rule before people rate anything. A pelican has to render, and a landing page has to be self-contained, work on desktop and phone, and include the required info.

Kaggle doesn't say on the page whether voting needs an account.