We have routed most of our side-project traffic through OpenRouter since April. Verdict up front: keep it for prototyping and model-shopping, measure it before production.
The good
One key, one OpenAI-compatible endpoint, essentially every model that matters. Switching a project from one model to another is a one-line change. The free-tier models (the ones suffixed :free) are real and cost nothing.
The meh
Latency adds up. Our measured overhead versus calling providers directly averaged 200–400ms per request. For a chat app you will not notice; for an agent making dozens of calls, you will.
The catch: free models are rate-limited to roughly 20 requests per minute and can be deprecated with little notice — we had one model disappear mid-project. And on paid models, OpenRouter adds a fee on top of provider pricing; check the pricing page for the current cut before committing real volume.
Who it is for
Tinkerers comparing models, developers who want a fallback chain, anyone tired of managing five API dashboards. Who it is not for: latency-sensitive production apps at scale, where direct provider contracts win.