Six months ago, the engineering lead at a 40-person SaaS company pulled up a spreadsheet in our call. It listed five separate LLM subscriptions: Anthropic for Claude, OpenAI for GPT, Google for Gemini, and two Chinese providers for GLM and KIMI. Five billing cycles, five API keys floating in various .env files, five dashboards to check when a developer said "the AI is slow today." Nobody on the IT team could answer a basic question: which IP addresses are we actually calling from, and which ones need to stay on the allowlist?
[..]
Read more...