Benchmarking is cool, but for production I care about real inference latency, self-hosting VRAM costs, and how cleanly it handles structured JSON output.