🛠️ What AI gateway do you run in production — and what broke?

[cuihuan]

One thing we keep learning: the best signal about a gateway isn't its star count — it's what actually broke for you in production. So let's pool it.

If you run an AI gateway / LLM proxy / router, share:

1. What you run — LiteLLM / OpenRouter / Portkey / Bifrost / Kong / new-api / self-rolled / …
2. Hosted or self-hosted, and roughly your scale (req/day)
3. What it does well — the reason you picked it
4. What broke — the 3am-pager moment: latency, memory leaks, prompt-cache not passing through, tool-calls mangled, an outage, a surprise bill…
5. Would you pick it again?

Real reports beat marketing. We fold recurring, verifiable findings into the scorecard and the production-reality notes (with credit). No vendor pitches please — this thread is for operators.