LLM gateway 工程笔记 — BYOK / prompt cache / 计费 / 可观测https://blog.synthorai.io/https://blog.synthorai.io/blog/what-is-an-llm-api-gateway/https://blog.synthorai.io/blog/what-is-an-llm-api-gateway/A technical introduction to LLM API gateways — what they do, when you need one, and how they differ from traditional API gateways. Covers architecture, BYOK, billing, and cross-protocol translation.Tue, 12 May 2026 00:00:00 GMThttps://blog.synthorai.io/blog/choosing-an-llm-gateway/https://blog.synthorai.io/blog/choosing-an-llm-gateway/The LLM gateway category has settled into three architectural shapes — managed SaaS, in-app library, and self-hosted service. A practical comparison of trade-offs, plus a decision matrix and migration paths.Tue, 12 May 2026 00:00:00 GMThttps://blog.synthorai.io/blog/byok-billing-invariants/https://blog.synthorai.io/blog/byok-billing-invariants/BYOK looks like a feature checkbox but is the largest class of correctness bugs in production LLM gateways. Five billing invariants we enforce by test, with concrete examples of what breaks when each is violated.Tue, 12 May 2026 00:00:00 GMThttps://blog.synthorai.io/blog/prompt-caching-across-providers/https://blog.synthorai.io/blog/prompt-caching-across-providers/Anthropic, OpenAI, and Gemini all support prompt caching with different APIs, TTLs, and pricing. A gateway has to translate between them — and some of the translation is lossy. The rules, the edge cases, and the tests that catch the bugs.Tue, 12 May 2026 00:00:00 GMT