tl;dr: Running production traffic on OpenRouter reveals that "same model, different provider" can mean 20-point benchmark swings, broken vision support, ignored reasoning-effort settings, unreliable tool-call parsing, and 200 OKs with empty content. Declared quantization (fp4/fp8) is a poor quality proxy, provider contracts differ on things like reasoning_content history, and rate limits often hit by IP so laptop testing lies. Pinning providers doesn't save you either—reliable ones go down, rate-limit, or silently stop serving models, so you need defensive parsing, retries, and per-workload benchmarking from prod.