I just run my own benchmark for "draw an SVG with $animal driving $vehicle". I won't post my choice of animal and mode of transport, but there are plenty of uncommon combinations to choose from. So far it's a fun and visually intuitive benchmark that does seem to correlate with model capabilities