What Should an AI Benchmark Measure?

A single score is easy to compare, but often hides the trade-offs that matter in practice. Should a benchmark emphasize average quality, worst-case errors, latency, cost, or performance across different user groups?

For generative systems, I find it useful to separate correctness, usefulness, and response time, then inspect examples where those goals conflict. What metrics or evaluation practices have helped you make better model choices?