I suggest benchmarking models quarterly using your own use cases; monitoring changes in model pricing, latency and feature ...