Token Goblin :[]: Field testing live voice + live data w/ GPT-Realtime-2.1 api // Managing Costs?

I've been daily field-testing Realtime-2.1/Live Transcribe + live Telemetry data (GPS/IMU), and it's damn impressive with tools/reasoning enabled, but the costs are like a hoard of level 10 sneaky goblins. A very active user can gobble up around $5/day in tokens. My field testing is around $8-10/day. Looking for advice or live usage insights on how costs were managed for production. -------- Usage examples: (hands-free co-pilot app) Live data requests Function/feature launching Requesting summaries/comparisons of telemetry data Real-time playback of urgent data Logging voice notes/expenses Time-based requests with actions Details: Model: gpt-realtime-2.1 (WebRTC) Reasoning: low Semantic VAD: Medium Auto response: On Auto interruption: off (tap-based UX) Noise reduction: far-field Live Transcribe: low-delay Tool Choice: auto (19+ tools) Automatic prompt caching Heavy contract compaction for tools Limited instructions/personas Tailored routing for token limits for response (320-1024 token limit range)