NotCheap MAKE EVERY CREDIT COUNT
Analyze Your Costs

GPT-Live-1 Pricing: Budget a Five-Minute Voice Call

Share this guide

Here's a misconception I see trip up a lot of teams building conversational voice bots. They spot OpenAI's published headline rate of $0.05/minute on the model card, multiply by five, and assume a customer call costs a clean quarter. In production, that quarter only opens the microphone. A five-minute GPT-Live-1 session costs $0.25 for the voice connection alone, before you add backend model tokens, tool execution, or telephony routing. In our realistic baseline scenario, the full workflow runs $342.00 for 1,000 completed calls with GPT-6 Luna, or $382.30 with GPT-6.1 Sol, once you account for real-world failed connection attempts. Think of those as operational project budgets, not a flat marketing quote.

OpenAI's official GPT-Live-1 model page prices audio streaming at $0.05/minute billed per second, with backend reasoning charged as a separate meter. All rates verified 25 October 2026; USD before sales tax or carrier surcharges.

Pay for the open session, including quiet time

GPT-Live is your real-time audio pipeline, streaming input and output audio with sub-second responsiveness. The heavy lifting (reasoning, business logic, API calls) happens behind the scenes through a connected backend intelligence model. As OpenAI's voice cost guide makes clear, session duration doesn't care whether anyone is talking. Dead silence, customer pauses, and hold music all burn the same clock. Muting your microphone in the browser or on a SIP trunk won't stop billing either; as long as the WebSocket or WebRTC channel stays open, the meter runs.

When you establish a WebRTC channel, session creation bills 15 seconds during initialization. The good news is that OpenAI credits this startup window against the active session's running clock, so a five-minute session is not five minutes plus another 15 seconds. The practical catch is that if an end user hangs up before the audio stream ever connects, that abandoned initialization still leaves a charge of $0.0125. That's why your production monitoring should track session creations and dropped handshakes right alongside completed conversations.

Define one completed call before multiplying it

Let's model a standard 300-second (five-minute) customer support inquiry. In our benchmark scenario, the caller speaks for roughly 90 seconds, the assistant speaks for 120 seconds, and thinking pauses, tool executions, and audio buffering fill the remaining 90 seconds.

Across ten backend turns, assume a cumulative total of 15,000 input text tokens and 1,200 output text tokens. Conversational context accumulates as the dialogue progresses, so that 15,000-token figure already covers the rolling system prompt and message history across all ten backend turns. No need to multiply it by ten again.

Using OpenAI's current text rate card, standard uncached backend reasoning breaks down like this:

LayerCalculation per callUsage charge
Live voice session5 minutes × $0.05$0.2500
GPT-6 Luna backend0.015M × $0.10 + 0.0012M × $0.50$0.0021
GPT-6.1 Sol backend0.015M × $2 + 0.0012M × $10$0.0420

Looking strictly at AI inference, your model-only totals come out to $0.2521 with Luna and $0.2920 with Sol per call. Luna is vastly cheaper on paper, but cost per token shouldn't be your only decision criterion. If Luna misreads account numbers or stumbles on structured tool arguments, the human escalation costs will quickly dwarf any API savings. I'd benchmark both models against real, messy caller recordings before you commit to a routing architecture.

Add external services as separate rows

An AI voice model can't dial a phone number on its own. For our production calculation, we add two operational layers: a third-party SIP trunking or PSTN telephony provider at $0.015/minute, plus an external customer verification or CRM database lookup at $0.01 per completed call. These are typical market rates, not a bundled OpenAI tariff, so swap in your contracted carrier rates.

Over five connected minutes, telephony adds $0.075, while the database query adds $0.01. That brings your true cost per completed call to $0.3371 with Luna, or $0.3770 with Sol. If you deploy entirely inside a web browser or mobile app, with no phone carrier and no billable API lookups, you can safely drop those two rows.

Completed calls, no failed attemptsVoice + backend onlyWith example phone + lookup
1,000, Luna backend$252.10$337.10
1,000, Sol backend$292.00$377.00
10,000, Luna backend$2,521.00$3,371.00
10,000, Sol backend$2,920.00$3,770.00

Keep in mind that these figures are raw utility spend. They leave out upfront agent engineering, latency monitoring tools, audio recording storage, human supervisor fallbacks, and platform support packages. It's cleaner to budget that infrastructure separately than to inflate your per-minute telephony number to cover it.

Failed attempts have a duration, not just a percentage

In any production telephony queue, not every connection ends in a happy resolution. Users hang up during the greeting, voicemail picks up, or a cellular connection drops in transit. For this model, we budget 100 failed or abandoned attempts for every 1,000 completed calls (a realistic 10% failure rate). Each failed attempt lasts an average of 45 seconds (0.75 minutes), processes roughly 1,500 backend input tokens and 120 output tokens, incurs telephony carrier charges, and ends before triggering any paid CRM lookup. Credited startup time is already absorbed within that 45-second duration.

Here's the math: a Luna failure costs 0.75 × $0.05 + 0.0015M × $0.10 + 0.00012M × $0.50 + 0.75 × $0.015 = $0.04896. A Sol failure, using the $2/$10 text token card, costs $0.05295.

Completed calls plus short failuresLuna full workflowSol full workflow
1,000 completed + 100 failed$342.00$382.30
10,000 completed + 1,000 failed$3,419.96$3,822.95

My advice is to calculate your forecast from unrounded per-call micro-cents and round only when you reach the final monthly invoice. Tacking an arbitrary 10% markup onto your completed-call total muddies the picture, because failed calls end early and skip the expensive backend tool executions. And if zero sessions run, your usage bill is $0.00, though dedicated phone numbers and reserved SIP trunks will still carry small recurring monthly fees.

Realtime audio tokens are a different meter

It's easy to mix up GPT-Live-1 with OpenAI's GPT-Realtime-2.1 API, so let's untangle them. Realtime-2.1 drops session-duration billing entirely in favor of direct audio token metering. OpenAI estimates approximately 600 input audio tokens per spoken minute and 1,200 output audio tokens per spoken minute, supplemented by text tokens for system prompts and conversation history. Speech cadence, background noise, and cached prefix hits all swing token generation quite a bit, so you can't accurately price a Realtime-2.1 deployment by multiplying call minutes by a flat output rate.

The voice cost guide walks through the technical mechanics and prompt caching behavior. Unless you've collected hundreds of hours of production call telemetry to verify your exact audio token ratios, GPT-Live-1 gives you much more predictable financial guardrails for contact centers.

Our recommendation: start with the transparent session-plus-backend formula, and set up strict automated hang-up rules for dead air. Trimming just 60 seconds of idle hold time saves $0.05 in OpenAI audio fees and $0.015 in telephony on every single call. Solid session lifecycle management will save your company far more than shaving pennies off backend token prompts.

PASS IT ON

Useful for your team?

Share this guide with a colleague.