Autorouters, the tools that draw the copper connections on a printed circuit board (PCB) once the parts are placed, have mostly been sold on screenshots and the vendor's own example boards. Koppar AI has published a benchmark that takes the scoring away from the router. Every tool gets the same placement, the same design rules, and the same time limit, and the open-source design tool KiCad checks what comes back: whether every connection is actually made, and whether the board passes design rule checking (DRC, the automated test that copper spacing and widths follow the manufacturing rules).
The first round uses eight real open-hardware boards, from 2 to 8 layers, with their original routing stripped out. A board only counts as finished if every connection is made and the router added no new rule errors. Under the published default settings, Koppar's own router finished all eight, TraceMaker five, Zap! three, and Freerouting, tscircuit, and KiCad Routing Tools none. The harness also records vias, track length, run time, ground plane coverage, how differential pairs and power rails were handled, and the paths to decoupling capacitors.
Koppar built the test and won it, so treat the ranking with care. What makes it worth using anyway is that everything is pinned: the source boards are checksummed, the scripts are public, and KiCad's result doesn't depend on who ran it. If you're choosing a router, add two of your own recent boards to the harness and see which tools still finish.