New Benchmark Reveals Top AI Models Still Struggle With Visual Math Reasoning