Good question it’s a bit tricky because the KB doesn’t follow the same per-minute model as LLM + transcription. The best way is usually to run a few sample calls and measure usage, then factor in retrieval + embedding costs alongside the per-minute pieces. Dm me