Antirion: Your AI gateway should add 0.5 ms, not 20

Hey Product Hunt 👋 I'm Elias, and I built Antirion. At my last job our LLM gateway was the slowest hop in the request: it added more latency than the network, ate gigabytes of memory doing nothing, and still couldn't tell me which team burned the budget. Every team shipping AI ends up needing the same thing: one place for provider keys, fallbacks when a provider goes down, budgets per team, guardrails, and a log of what every call cost. I wanted all of that without paying for it in latency. So Antirion is one Rust binary, with the console built in: ⚡ 0.53 ms of overhead at 5,000 RPS on 4 vCPUs. Bifrost adds 19.7 ms (37x slower) on the same machine, and LiteLLM fails at that rate. 🔀 OpenAI, Anthropic, Bedrock, Vertex, Gemini, Mistral, vLLM and more behind one API, with fallbacks, retries, hedging and circuit breakers. 💸 Budgets and rate limits in dollars or tokens per org, team, project and key, with alerts before they run out. 🛡️ PII and secrets redaction and prompt-injection checks, even on streams. 🔎 A request log that shows the route, every retry, the tokens and the cost of each call. Don't trust the numbers? The benchmark harness is in the repo, so you can rerun it on your own AWS account. It's Apache-2.0 and fully self-hosted. One question for you: what's the first thing you'd route through it, and what would stop you from putting it in production? I'll be here all day answering 🙏