OpenAI Releases GPT-Live-1: Realtime Full-Duplex Voice API, $0.05/Min Economics, and Sub-300ms Turn-Taking

OpenAI launched GPT-Live-1 in its developer API, opening the full-duplex conversational voice system previously restricted to ChatGPT Advanced Voice Mode. The release provides developers with a dedicated, low-latency foundation model designed for bidirectional audio streaming.

GPT-Live-1 processes acoustic tokens natively. By eliminating the traditional three-hop pipeline of speech recognition, text inference, and speech synthesis, the model achieves median end-to-end audio latencies of 280ms to 320ms. It handles conversational interruptions, filters background kitchen and street noise, and executes dynamic handoffs to reasoning models for heavy compute tasks.

OpenAI priced the voice layer at $0.05 per minute ($3.00 per active hour). Enterprise adopters, including Yelp and Hatch, deployed the model across customer-facing phone networks and consumer hardware following pilot tests exceeding one million real-world calls.

---

The Latency Problem in Cascaded Voice Stacks

Most commercial voice bots operate on a cascaded architecture:

[User Speech] 
      │ (Audio Stream)
      ▼
1. Automatic Speech Recognition (Whisper / Deepgram) ──> [Text Transcript: 250ms - 450ms]
      │
      ▼
2. LLM Text Inference (GPT-4o / Claude 3.5) ───────────> [First Token Generation: 300ms - 600ms]
      │
      ▼
3. Neural Text-to-Speech (ElevenLabs / Cartesia) ──────> [Audio Buffer & Chunking: 250ms - 400ms]
      │
      ▼
[Bot Voice Output] ──> Total Latency: 800ms - 1,450ms

This cascade creates three distinct operational bottlenecks:

1. Compounding Latency: Each stage buffers data before passing it to the next. Even with token-chunked streaming, total round-trip time regularly exceeds 900ms. Human conversation breaks down when turn-taking latency passes 400ms.
2. Loss of Non-Verbal Acoustic Signal: Transcribing audio into text strips away pitch, sarcasm, hesitation, volume modulation, emotional cadence, and ambient acoustics. The LLM receives flat characters and responds with generic prose.
3. Unnatural Interruption Mechanics: In a cascaded system, handling user interruptions requires external Voice Activity Detection (VAD) heuristics. The system frequently cuts off speakers when they pause to think, or fails to stop talking when a customer speaks over it.

GPT-Live-1 replaces this entire cascade with a single, end-to-end multimodal transformer. Audio enters the model as continuous acoustic tokens, and speech returns directly as 24 kHz PCM audio packets.

---

Architectural Mechanics: Full-Duplex Acoustic Tokens

+-------------------------------------------------------------------------+
|                  OpenAI GPT-Live-1 End-to-End Pipeline                  |
+-------------------------------------------------------------------------+
|                                                                         |
|  [Microphone Audio (WebRTC / Opus)]                                      |
|                 │                                                       |
|                 ▼                                                       |
|   [Acoustic Neural Codec Encoder] (12.5 Hz frame rate)                   |
|                 │                                                       |
|                 ▼                                                       |
|   [GPT-Live-1 Autoregressive Core] <──────> [Context & Session State]   |
|         │                        │                                      |
|         │ (Simultaneous Input)   │ (Interruption Detector: <80ms)       |
|         ▼                        ▼                                      |
|   [Neural Audio Decoder]   [Asynchronous Tool Dispatcher]                |
|         │                                │                              |
|         ▼                                ▼                              |
|  [24 kHz PCM Audio]              [Delegation to o3-mini / GPT-4o]       |
|                                                                         |
+-------------------------------------------------------------------------+

GPT-Live-1 uses a unified cross-attention architecture that computes speech recognition, reasoning, and speech synthesis within the same parameter space.

Key architectural properties include:

• Full-Duplex Listening: The model ingests audio tokens on its input stream while emitting audio tokens on its output stream. It does not toggle between "listening mode" and "speaking mode."
• Acoustic Interruption Detection: Because the network processes user audio during its own generation steps, it detects semantic intent within 80ms of human speech. If the user utters an acknowledgement ("uh-huh", "right"), the model continues speaking. If the user issues a contradiction ("wait, stop there"), the model immediately truncates output generation and updates its internal state.
• Background Noise Isolation: The audio encoder distinguishes primary speaker formant frequencies from secondary background babble, dishes clattering, and siren wails without relying on separate audio-cleaning preprocessing layers.

---

Empirical Benchmark Performance

OpenAI evaluated GPT-Live-1 against commercial voice stacks on conversational responsiveness, barge-in accuracy, and subjective naturalness.

System Pipeline	Turn-Taking Latency (P50)	Interactivity Score (%)	Barge-In Precision (%)	Cost per Audio Hour
Cascaded Stack (Deepgram + GPT-4o + ElevenLabs)	880ms	52.4%	61.2%	$7.20 – $14.40
Cascaded Fast Stack (Deepgram + Groq Llama-3 + Cartesia)	460ms	64.8%	71.0%	$4.80 – $6.50
OpenAI gpt-4o-realtime-preview (Oct 2024)	440ms	69.5%	74.3%	$18.00 (Text + Audio tokens)
OpenAI GPT-Live-1 (Production Release)	290ms	80.1%	88.6%	$3.00 ($0.05/min flat)

The Interactivity Score measures how smoothly a voice agent negotiates conversational boundaries: handing off the floor when finished, holding the floor during natural cognitive pauses, and yielding the floor upon user interruption without clipping or stuttering. GPT-Live-1 scored 80.1%, outperforming cascaded pipelines by 27.7 percentage points.

---

API Economics: $0.05 per Minute

Earlier real-time voice APIs proved prohibitive for mass-market telephone workflows. In initial preview builds, high per-token audio rates pushed hourly operational costs past $15.00, restricting deployment to high-margin niches.

OpenAI structured GPT-Live-1 pricing around a flat rate:

$C_{ ext{minute}} = $0.05 quad implies quad C_{ ext{hour}} = $3.00$

This represents an 83% price reduction compared to earlier real-time preview iterations. For enterprise contact centers and consumer hardware, the flat $0.05/min rate provides predictable margin calculations:

• 10-Minute Support Call: $0.50
• 30-Minute Tutoring Session: $1.50
• 10,000 Inbound Inquiries (3 mins average): $1,500.00

Traditional multi-vendor stacks require separate API keys, three distinct network round-trips, and billing across speech recognition, inference, and voice cloning providers. GPT-Live-1 unifies these layers into a single WebRTC socket connection.

---

Architectural Pattern: The Model Handoff

A voice model running at 300ms latency cannot also execute 30-second chain-of-thought mathematical proofs or parse 50,000-line database schemas. Running frontier reasoning on every spoken phoneme is computationally inefficient.

OpenAI designed GPT-Live-1 to operate as an acoustic interface layer that executes model handoffs.

[User Speaks: "Can you rebook my flight to Denver for Thursday morning?"]
                     │
                     ▼
             [GPT-Live-1 Core]
                     │
                     ├─► Emits Filler Audio: "Looking up Thursday morning flights to Denver right now..."
                     │   (Keeps conversation active, 0 dead air)
                     │
                     └─► Dispatches Asynchronous Tool Call
                                 │
                                 ▼
                     [Background Reasoning Model: o3-mini / GPT-4o]
                                 │
                                 ├─► Queries Airline Reservation API
                                 ├─► Resolves Seating Preferences & Fare Rules
                                 └─► Returns Structured JSON Payload
                                 │
                                 ▼
             [GPT-Live-1 Resumes Voice Output]
                     │
                     ▼
[Bot Speaks: "I found a nonstop flight leaving at 8:15 AM on United. Would you like me to confirm seat 12B?"]

During asynchronous tool execution, GPT-Live-1 produces conversational acoustic bridges ("Let me check that inventory...", "One moment while I open your reservation..."). This prevents the dead silence that causes users to hang up or repeat themselves during complex backend operations.

---

Enterprise Deployments: Yelp and Hatch

Production deployments validate how models handle real acoustic environments. Two commercial platforms led testing through initial rollouts.

Yelp: 1,000,000 Local Business Calls

Yelp integrated GPT-Live-1 into its inbound call answering system for local service providers, including plumbers, auto repair shops, and restaurants. These businesses frequently miss customer calls when staff are actively working on job sites or managing physical kitchens.

During tests covering more than one million real telephone calls, Yelp measured three critical operational outcomes:

• Ambient Noise Resistance: Callers phoning from noisy roads or echoing garages experienced zero false interruptions. The model sustained coherent dialogue without getting derailed by car horns or background radio music.
• Mid-Sentence Revisions: When callers changed their intent midway through a request ("Actually, wait, don't book Tuesday, make it Wednesday afternoon"), GPT-Live-1 discarded the prior booking slot and updated the reservation form accurately.
• Call Completion Rates: Automated appointment booking completion increased by 34% compared to legacy touch-tone and cascaded interactive voice response (IVR) platforms.

Hatch: Calming Audio and Sleep Coaching

Hatch deployed GPT-Live-1 into its connected smart sleep devices to deliver conversational bedtime routines, personalized children's stories, and guided relaxation exercises.

Unlike contact center bots that demand rapid information density, sleep routines require:

• Soft, whispering vocal timbres and relaxed cadence.
• Dynamic attenuation: lowering volume and pacing speech progressively slower as the listener grows drowsy.
• Zero harsh vocal artifacts or sudden robotic pitch swings.

GPT-Live-1 adjusted its prosody based on conversational context, maintaining a soothing, quiet delivery without degrading speech intelligibility.

---

Developer Integration: LiveKit WebRTC Agent

The fastest way to deploy GPT-Live-1 into production is via LiveKit, the open-source WebRTC infrastructure network. LiveKit published direct support for GPT-Live-1 via its Python and TypeScript agent frameworks.

The following TypeScript implementation establishes an authenticated, bidirectional WebRTC session, defines a structured function tool, and configures acoustic interruptions:

import { VoicePipelineAgent, JobContext, cli } from '@livekit/agents';
import * as openai from '@livekit/agents-plugin-openai';

export default cli.defineAgent({
  entry: async (ctx: JobContext) => {
    // 1. Connect to the room over low-latency WebRTC
    await ctx.connect();
    console.log('[Agent Connected] Room: ' + ctx.room.name);

    // 2. Initialize GPT-Live-1 via the Realtime API plugin
    const liveModel = new openai.realtime.RealtimeModel({
      model: 'gpt-live-1',
      voice: 'alloy',
      temperature: 0.7,
      instructions: 'You are an expert customer concierge for Alpine Peak Logistics. Speak concisely, warmly, and naturally. If you need to query database records, inform the user you are looking it up.',
      turnDetection: {
        type: 'server_vad',
        threshold: 0.5,
        prefixPaddingMs: 300,
        silenceDurationMs: 200,
      }
    });

    // 3. Define structured business tools
    const checkTrackingTool = {
      name: 'checkShipmentStatus',
      description: 'Look up real-time delivery status for a tracking number',
      parameters: {
        type: 'object',
        properties: {
          trackingId: { type: 'string', description: 'The 10-digit shipment tracking code' }
        },
        required: ['trackingId']
      },
      execute: async ({ trackingId }: { trackingId: string }) => {
        // Simulating database lookup
        return {
          trackingId,
          status: 'Out for Delivery',
          estimatedArrival: 'Today by 3:30 PM',
          location: 'Denver Distribution Center'
        };
      }
    };

    // 4. Construct the voice agent pipeline
    const agent = new VoicePipelineAgent({
      llm: liveModel,
      tools: [checkTrackingTool]
    });

    // 5. Start listening to participant audio tracks
    agent.start(ctx.room);
    
    // 6. Greet the customer immediately upon entry
    await agent.say('Hello, thanks for calling Alpine Peak Logistics. How can I help you today?');
  }
});

---

Technical Considerations and Current Constraints

While GPT-Live-1 solves the core acoustic latency and interruption challenges of conversational AI, teams deploying the model must account for several operational realities:

1. Client Audio Echo Cancellation: Full-duplex speech-to-speech requires proper Acoustic Echo Cancellation (AEC). If the user's loudspeaker feeds audio back into their microphone, the model may interpret its own voice as an interruption. Hardware-level or browser-level AEC is mandatory.
2. Context Window Limitations: Long phone calls accumulate audio and text tokens rapidly. For conversations extending beyond 30 minutes, engineering teams must implement rolling context truncation or summarize prior conversational turns into system prompt memory.
3. Regional Telecom Jitter: In telephone integrations (PSTN/SIP), packet loss and jitter over cellular networks can degrade audio frames. Deploying WebRTC edge relays close to telecom carriers preserves sub-300ms performance.

OpenAI's launch of GPT-Live-1 shifts conversational software from disjointed turn-taking scripts into continuous, fluid spoken dialogue. At $0.05 per minute, voice agents can now operate at commercial software scale.