Today I had fun playing with 𝗷𝗲𝘃 and routing it via 𝗢𝗽𝗲𝗻𝗥𝗼𝘂𝘁𝗲𝗿 + 𝗢𝗽𝗽𝗲𝗿 𝗔𝗜 in my homelab n8n. I optimized two workflows where I used OpenAI 5.6 Luna in a decision process — exactly the kind of “fast decision” case jev is made for. Cheaper, much faster, and structured output without needing a separate parser. I still didn’t manage to get it down to 1¢ per run, though. Lots more ideas where jev can make LLM-based processes cleaner. The AI space is moving insanely fast.