Latency, throughput, concurrency and cost for every model + engine setting we served — one sheet per model. All throughput is aggregate across concurrent streams; per-stream is shown next to it.