Todoist Ramble currently runs on Gemini Live 2.5 Flash. Across 197 live-audio tests covering 31 language variants, GPT Realtime 2.1 Mini raised our pass rate from 33% to 75% while cutting model costs by 28%! This is 2x+ the performance, for less money. Meanwhile, Gemini Live 3.1 still isn’t available on Vertex, and the split between the Gemini API and Vertex AI remains confusing. It’s really unclear what Google is doing, but from a dev perspective, it looks like a mess where they are falling…