I'm on Claude Max 20x ($200/mo) which gives me plenty of headroom, but you definitely don't need that. A 5x plan works fine, or you can go the API route through OpenRouter and pay per token. The voice layer (Parakeet + Kokoro) runs locally on the M4 so that's zero cost. The key is good memory practices so the agent doesn't waste tokens re-learning things, and falling back to cheaper models (like Haiku or Gemini Flash) for simple tasks. The heavy model only kicks in when it actually needs to think.