GPT-Live API vs. Realtime API: one voice frontend, one voice model

GPT-Live talks with the user and hands the thinking to a backend you choose. Realtime does the listening, reasoning, tool calls and speaking in one model session. The two bill on different units.

Duplex Compute · October 13, 2026 · Rates and versions checked October 13, 2026

gpt-live-1 is a full-duplex voice model. It listens and speaks at the same time, and it delegates reasoning and tool use to a backend agent. OpenAI calls that hand-off delegation. You pick the backend model separately from the voice model.

The Realtime API puts speech, reasoning and tool use in one session. One model interprets the audio, decides what to do and answers in speech. Its current model is gpt-realtime-2.1, a reasoning model with function calling. It handles barge-in: with voice activity detection on, the API detects user speech, cancels the ongoing response and starts a new one.

Both are speech APIs from the same company, with different shapes. GPT-Live is a conversational frontend for an agent you already run. Realtime is the agent.

Sources for the table: the GPT-Live 1 model page and GPT-Realtime-2.1 model page for endpoints and modalities; the GPT-Live and Realtime telephony guides for SIP; the Realtime conversations guide for the session cap.

Client delegation and Responses delegation

GPT-Live has two ways to run its backend, and you choose one when you create the session.

With client delegation, your application connects its own agent, on any model or provider, runs the work and sends results back to GPT-Live. This is the route for a team that already has a text agent with business rules and tools. The voice layer goes in front of it and the agent stays where it is.

With Responses delegation, OpenAI runs a supported Responses model as the backend and passes context and results between it and GPT-Live. Your application still runs its own function tools. To switch modes you start a new session.

OpenAI's guidance splits the prompts. GPT-Live gets a short prompt about speaking style and when to delegate. The backend gets the detailed instructions, business rules and tool workflows. In both modes your application checks permissions, collects confirmations and saves task progress.

Interrupting the voice does not cancel the backend

GPT-Live keeps talking with the user while backend work runs. The user can add a detail or change their mind mid-lookup. OpenAI's guide states the consequence plainly: backend work can continue when the caller interrupts the assistant, and your application decides whether to finish or cancel it.

t=0.0s  user:     "Book the 7pm table for six."
t=1.2s  GPT-Live: delegates booking to backend
t=1.3s  backend:  starts reservation call
t=1.5s  GPT-Live: "Checking that for you..."
t=2.4s  user:     "Actually, seven people."
t=2.4s  GPT-Live: stops speaking, listens
t=3.1s  backend:  reservation for six still running
        app:      must cancel it, amend it, or let it finish and correct after

The times above are illustrative, not measured. The point is the last line. Stopping speech is a voice event. Cancelling a reservation is an application decision. If your backend writes to a database or charges a card, give each action an idempotency key and decide in code what a correction does to work already in flight.

Realtime has a related problem in a different place. The model makes the tool call itself, then your code runs the function and returns a function_call_output item with the matching call_id. An interruption cancels the model's response. Your function is still your code, and what happens to its side effects is still your choice.

What 10 minutes costs on each

The two products meter different things, so the same conversation produces two very different calculations. Both examples below are illustrative arithmetic from published rates, not invoices. Rates checked October 13, 2026.

GPT-Live: session seconds

GPT-Live voice sessions cost $0.05 per minute, billed per second and not rounded up. Active session time includes the user speaking, the assistant speaking, both being silent and the backend working. Muting the microphone does not stop the meter. Closing the session does.

An hour of open session is $3.00 before backend costs. Creating a WebRTC session bills 15 seconds during initialization, which is credited against the session once it runs. Do not add those 15 seconds to every successful session. For reconciliation, record the final usage.seconds from session.closed. Usage updates are cumulative snapshots, so summing them overcounts.

Realtime: tokens by modality

gpt-realtime-2.1 charges $32 per million audio input tokens, $0.40 per million cached audio input tokens and $64 per million audio output tokens. Text is $4 input, $0.40 cached and $24 output per million. User audio is 1 token per 100 ms and assistant audio is 1 token per 50 ms, so a minute is about 600 and 1,200 tokens respectively.

The $0.384 is a floor for this example. Realtime sends the whole conversation to the model on each response, so later turns cost more. Cached input is much cheaper, and caching is best-effort. Input transcription, if enabled, runs on a separate model with its own rate card. Silence costs less here than on GPT-Live: VAD filters empty input audio, so quiet stretches don't become input tokens.

Neither number is the full price of a task. GPT-Live's bill grows with wall-clock time. Realtime's grows with speech, turns and history. Our voice agent pricing note has the formulas for both.

Session limits and concurrency

A Realtime session can last at most 60 minutes. That cap belongs to Realtime. The GPT-Live session guide lists an expired close reason for a session that reached its duration limit without giving a number. Plan reconnects for both. GPT-Live's guides describe seeding a new session with saved context and the verified task result.

GPT-Live rate limits are counted in concurrent sessions: 50 on the Build tier, 300 on Launch and 500 on Grow. Realtime limits are listed as requests and tokens per minute on its model page.

When to choose which

OpenAI publishes no head-to-head latency comparison of the two, and we haven't run one. If speed decides it, measure the time to a useful spoken answer on your own scenarios, as the voice agents guide describes, and keep "I'm checking" fillers separate from the answer.

Duplex Compute is another full-duplex option: a voice API with tool calling and long conversations, billed at $0.01 per connected minute. Ten minutes is $0.10 of voice-session rate. As with GPT-Live, that rate covers the voice session, so compare complete task costs before deciding.

Related notes

• Voice agent pricing: calculator and billing formulas
• OpenAI Realtime function calling
• Why your Realtime agent keeps talking after an interruption
• OpenAI Realtime API alternatives
• Full-duplex voice AI vs. turn-based