I Priced a GPT-Live-1.1 Voice Receptionist: $140 per 1,000 Calls (and Where the Hidden Bill Lives) A developer at VisionQuantech priced OpenAI's GPT-Live-1.1 full-duplex voice model for an AI phone receptionist, finding it costs $140 per 1,000 four-minute calls with 50% silence, versus $200 all-active, thanks to a new $0.02/min standby rate that is 60% cheaper than the $0.05/min active rate. The analysis warns that the voice layer is only part of the bill, since backend model and tool calls are billed separately at normal rates and are not discounted during standby, and that standby is WebRTC-only and offers little savings on short, dense calls. By Shivam Kumar, VisionQuantech — 22 November 2026 OpenAI shipped GPT-Live-1.1 gpt-live-1.1 to the API on 20 November 2026, and as a builder the first thing I did was price it. Not the demo. The real thing: an AI receptionist answering a business phone line all day. The headline number is good — $140 per 1,000 calls — but the bill has a second half that nobody quotes. Let us do the math properly. GPT-Live-1.1 is a full-duplex voice model: it listens while it speaks, so callers can interrupt. Sessions start at the v1/live/sessions endpoint. Version 1.0 went GA on 10 September 2026 at $0.05/min; 1.1 keeps that active rate and improves transcription accuracy and accent/pronunciation stability across languages. The real change is standby pricing. Active conversation costs $0.05/min. Standby — the line open, nobody talking — costs $0.02/min, 60% cheaper. Billing is per second. On standby the model's inference pauses but the WebRTC connection, session ID, and conversation history stay alive, and speech wakes it. The API fires session.paused and session.resumed events so your backend can track state. Standby is WebRTC-only, and you configure it on the session: { "wake policy": { "type": "speech" }, "idle timeout s": 15, "vad threshold": 0.5 } idle timeout s takes 5–60 seconds: how long a quiet line waits before dropping into standby. vad threshold runs 0–1 default 0.5 and tunes voice-activity-detection sensitivity — lower catches quieter speech, higher ignores background noise. Pick the timeout to match your call pattern: a reception desk with long lookups wants a short timeout; a fast Q&A line wants a longer one so it does not keep pausing mid-sentence. There is also a 15-second WebRTC initialisation charge, credited once the session runs. Treat it as a setup cost, not a trap. Here is the scenario I priced: a small-business AI receptionist, line open 8 hours 480 minutes a day, 40% active talk and 60% standby — 192 active minutes, 288 standby minutes. That is $8.64 saved per day, 36% cheaper, about $190/month over 22 working days. Per-call, the picture is the same. Take 1,000 calls of 4 minutes each with 50% silence: 2,000 active minutes $100 plus 2,000 standby minutes $40 = $140 per 1,000 calls , versus $200 all-active. You save $60 per 1,000 calls — and the saving scales with silence, not with cleverness. Now the part that matters. The voice layer is not the whole bill. The backend model and every tool call your agent makes are billed separately at their normal rates, and standby does not discount them. If your receptionist calls a large reasoning model on every turn, or hits three tools to check a calendar, that token bill can dwarf the $140 voice cost. Standby saves you money on the voice layer only. Two more things to check before you ship. First, standby is WebRTC-only — if your telephony runs over SIP, verify your provider's bridge into WebRTC, because the discount does not exist on a pure SIP path. Second, standby barely helps short, dense calls: a two-minute rapid-fire exchange has almost no silence to discount. It pays off on long, hold-heavy calls — reception desks, clinics, garages — where the caller spends half the call waiting. Outbound SIP calling reached all organisations on 2 November 2026, so phone agents are a supported pattern now, not a hack. And if you are comparing models: GPT-Realtime-2.1 bills per token $32 in / $64 out per 1M audio tokens , while GPT-Live bills by duration. Dense short calls may favour tokens; long human-paced calls favour duration with standby. Bottom line for builders: price the full stack — voice layer plus backend model plus tools — on your real call pattern. GPT-Live-1.1 makes the voice layer genuinely cheap for hold-heavy lines. Just do not quote the $140 without the backend bill next to it. Sources: OpenAI model page — developers.openai.com/api/docs/models/gpt-live-1.1; MarkTechPost, 21 November 2026; APIdog guide to GPT-Live-1.1.