Benchmark Babel: Enterprise Buyers Face Evaluation Paralysis

Three new multimodal benchmarks (ARK, M-JudgeBench, UNICBench) arrived in 3 weeks using incompatible metrics. GPT-5.4 reports accuracy as percentages, Sonnet 4.6 as Elo ratings on the same benchmark. M-JudgeBench found 7B judge models are fundamentally unreliable. Enterprise teams lack the cost-per-task metric that actually matters.