GPT-Live-1 API pricing and cost per call

Pricing verified 11 September 2026 against the OpenAI pricing page and the gpt-live-1 model page. OpenAI changes prices without changing this page, so check the source before you commit a budget to it.

GPT-Live-1 costs $0.05 per minute of voice session, billed per second, with no rounding up to a whole minute. That price buys the full-duplex speech layer and nothing else — the backend model and any tools it calls are billed separately at their own rates. There is no free tier, and rate limits are counted in concurrent sessions rather than requests per minute.

What you actually pay for on a call

A phone call through a GPT-Live-1 agent has three separate meters running, and only the first one is the $0.05 figure.

1. The voice layer. $0.05 per minute of session, billed per second. Session duration counts everything between start and close — the user speaking, the assistant speaking, both silent, and the backend working. Muting the microphone does not stop the meter; closing the session does. Creating a WebRTC session also bills 15 seconds up front while it initialises, credited back against the running session, so a session created before the caller is ready still costs those seconds.
2. The backend model and tools. Billed at normal rates for whichever model you delegate to, plus any tool or retrieval charges. This is the part the calculator cannot guess for you, because it depends on how much reasoning each call triggers and how long your prompts are. Measure it on real calls rather than estimating from a token count.
3. The phone line. Your carrier or SIP provider bills per-minute telephony separately from OpenAI, and those rates depend on country, number type and provider. We do not publish those figures here because we cannot verify them for your route — take them from your own provider's price list and add them to both columns of any comparison.

Full-duplex versus a cascaded stack

The alternative to a single full-duplex model is a cascaded pipeline: speech-to-text, then a language model, then text-to-speech, wired together by your own code. Its cost is three line items instead of one, each billed by its own unit — transcription usually per audio minute, the language model per token, speech synthesis per character or per minute. That is why the comparison mode in the calculator asks you for three per-minute figures: it is the only way to get the two architectures onto the same axis.

The comparison is not purely financial. A cascaded stack lets you inspect and rewrite the text at each stage, which is exactly what you need when a confirmation code has to be read back correctly. A full-duplex model speaks directly from audio, so there is no intermediate text for you to correct. Which matters more depends on whether the call is a conversation or a readout.

Concurrency, not requests per minute

GPT-Live-1 is rate limited by how many sessions are open at once, which is the natural unit for phone traffic — a call holds a session for its whole duration whether anyone is speaking or not. The published ceilings are:

Free is listed as an unsupported usage tier, so a paid tier is a prerequisite rather than an upgrade. Source: the GPT-Live-1 model page, read 11 September 2026.

The calculator's concurrency figure is traffic intensity: peak calls per hour multiplied by average call length in hours. Fifty calls an hour at three minutes each averages 2.5 simultaneous calls. Averages are not ceilings — calls arrive in bursts, and a tier limit is a hard wall that rejects sessions rather than queueing them, so size against your busiest minute and not your busiest hour.

Where IVR Voice Gen fits

Not every second of a phone call is a conversation. A main menu, an out-of-hours message, a hold announcement, a fixed legal disclaimer, a verification code read back at the end of a flow — these are the same words on every call. Generating that audio once and playing the file costs nothing per call, plays back identically every time, and can be signed off by whoever needs to approve the wording.

IVR Voice Gen is text-to-speech built for exactly those prompts: digit strings, reference codes, dates and email addresses read the way a caller can write them down. It is not a GPT-Live-1 integration and does not claim to be one. The two sit at different points in the same phone system — the hybrid architecture guide sets out which part belongs where, and the verification code use case shows what the pre-generated half looks like in practice.

Questions

Sources

• OpenAI: GPT-Live 1 model page
• OpenAI: Getting started with GPT-Live
• OpenAI: Voice latency and cost
• OpenAI: API pricing
• OpenAI: Telephony and SIP
• OpenAI: Voice latency and cost

All figures on this page were read from those pages on 11 September 2026. GPT-Live-1 has been available in the OpenAI API since 10 September 2026; it is served from the v1/live/sessions endpoint, and the older v1/realtime endpoints do not accept it. For contrast, GPT-Realtime-2.1 is billed per token rather than per minute ($32.00 per million audio input tokens, $64.00 per million audio output tokens), which is why the two cannot be compared on price without modelling a specific conversation.

IVR Voice Gen is not affiliated with OpenAI. GPT-Live is a trademark of OpenAI. This site does not integrate GPT-Live-1; the pages in this section are independent analysis based on the public documentation linked above.