OpenAI adds GPT-Live-1 real-time full-duplex VOICE model to API (discovery said "video" — CORRECTED)

On Thursday Sep 10, 2026, OpenAI announced and shipped, effective immediately, general availability of GPT-Live-1 in the API:

1. The model: GPT-Live-1 is OpenAI's premier full-duplex real-time voice model ("Our premier model for natural, expressive voice conversations with smooth interruption handling" — model card). It reasons over incoming and outgoing audio together in a single model, avoiding the latency and brittle handoffs of chained STT→LLM→TTS pipelines. It can listen while speaking, respond to interruptions and short acknowledgements ("mhmm", "yeah"), and keep a conversation running while a backend model or agent performs reasoning, tool calls, web search, or coding.
2. Delegation architecture (the defining feature): GPT-Live-1 handles conversation only; reasoning/tools live in a backend the developer chooses. Two modes: Responses delegation (OpenAI-managed — you configure a Responses model such as gpt-5.6-luna, gpt-5.6-terra or gpt-6-astra with tools like web_search or custom functions) and client delegation (your application builds the request and runs its own backend model, agent harness, or service; results are fed back via session.commentary.append / session.thinking.append). The launch post demonstrates pairing GPT-Live-1 with the Codex SDK for repository work mid-conversation. OpenAI's guidance is explicit: business logic, authorization checks, tool handlers and durable state stay in the application.
3. Transports: WebRTC for browser/mobile voice sessions (microphone/speaker media tracks + data channel), WebSockets for server-side audio streams, sideband connections for server control/observation of a live session, and Telephony/SIP support for phone agents (restaurant reservations, customer support, order updates).
4. New voice options: 12 new real-time voices (Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, Cinder) spanning accents, dialects and languages, plus the existing default marin; custom voices via sales eligibility.
5. Built-in speech plumbing: native ASR transcripts (session.input_transcript.delta) and response text (session.output_transcript.delta), strong alphanumeric understanding, keyword biasing, and native turn detection for developers who want explicit turn boundaries even though the model is not turn-based.
6. Pricing & limits: voice layer $0.05/min billed per second (not rounded up; a 90-second session = $0.075); backend model/tool usage billed separately at normal rates; free API tier not supported; rate limits measured in concurrent sessions: 25 (Tier 1) → 50 (Tier 2) → 200 (Tier 3) → 300 (Tier 4) → 500 (Tier 5). Knowledge cutoff: July 31, 2025.
7. Paired same-day launch: OpenAI also moved the Agents API into public beta the same day (Sep 10) — GPT-Live-1 is the voice that talks to a customer; the Agents API (managed Codex harness) is a backend it can hand the real work to. OpenAI additionally pointed enterprises at OpenAI Presence (introduced Jul 22) as the managed enterprise voice-agent product built on GPT-Live-1.
8. Day-one ecosystem: Twilio announced a native GPTLiveProvider in Agent Connect connecting Twilio Voice Media Streams directly to GPT-Live-1 sessions (Sep 10). Customer testimonials in the launch post: Yelp (Host/Hatch call handling), Speak (Live Tutor Lessons — ~80% fewer interruptions during thinking pauses), Intercom Fin (natural call flow for support), Cognition (Devin voice collaboration).

Context: GPT-Live-1 first shipped as the ChatGPT Voice model on July 8, 2026 (paid plans; GPT-Live-1 mini for Free). The API release on Sep 10 is the developer-surface debut of the same model family, 64 days later. OpenAI claims, in its own evaluations, a +30 percentage-point improvement on Full Duplex Bench over GPT-Realtime-2.1 and a #1 Tau3 finish when paired with GPT-6 Astra at medium reasoning effort (both COMPANY CLAIM).

• Before: Developers building real-time voice agents used the Realtime API (v1/realtime with gpt-realtime-2.1 / 2.1-mini, released Jul 6, 2026), a single all-in-one speech-to-speech model that handles audio, reasoning and tool selection inside one model. GPT-Live-1 existed only as a consumer ChatGPT feature (Jul 8) with plan-based hourly caps, no public model ID, no pricing, and a "notify me" signup for API access (as documented by multiple builders through mid-July).
• Change: OpenAI made GPT-Live-1 generally available to all developers through the new v1/live/sessions endpoint at $0.05/min (voice layer, per second), with full-duplex conversation handled by a dedicated voice model and reasoning/tools delegated to a separately chosen backend (Responses delegation or client delegation). SynthID watermarking had already been added to supported GPT-Live audio on Jul 31, signaling API readiness.
• After: Any developer with a paid OpenAI API key can now embed ChatGPT-Voice-grade full-duplex conversation in their own app, phone line or agent, choosing their own reasoning backend per task (cheap model for scheduling, frontier model for hard cases). The voice layer is commoditized at a legible price; the architecture splits "voice" from "brain," making each independently replaceable. Real-time voice moves from consumer novelty to a standard API building block — in the same week Google (Gemini 3.8 Live, Sep 15) and xAI (Voice Agent API at $0.08/min, Sep 17) shipped competing real-time voice surfaces.

• Single-model full duplex: GPT-Live-1 jointly models user input audio and generated output audio, so it can decide to keep talking, stop, acknowledge ("mhmm"), or pivot mid-stream without a voice-activity-detection state machine owned by the developer. This replaces the STT→LLM→TTS concatenation where each handoff adds latency and timing drift (OpenAI's framing, echoed by GIGAZINE's analysis).
• Session lifecycle (v1/live/sessions): WebSocket clients connect to wss://api.openai.com/v1/live/sessions and send session.start (model, instructions, audio format, voice, delegation config); the server replies session.started. WebRTC sessions are created with POST /v1/live/sessions (server exchanges the browser's SDP offer for an answer; audio travels on media tracks, JSON events on a data channel). Sessions are fixed-config at startup: model, instructions, voice and audio format are immutable (instructions can be appended via session.instructions.append; delegation settings are sparse-updatable within a mode).
• Delegation flow: GPT-Live-1 decides when a request needs the backend; it emits session.delegation.created with a delegation_id. In Responses delegation, it prepares a request to your configured Responses model (with tools such as web_search or functions), which runs and returns results through response.event envelopes; the voice model then communicates the result. In client delegation, your application assembles the backend call itself and returns content via session.commentary.append (content to speak, paraphrasable) or session.thinking.append (quiet context). Interrupting speech does not cancel backend work — the application controls permissions, task state and which results reach the conversation.
• Context management: instructions up to 16,384 tokens at startup; automatic conversation summarization/compaction — when context exceeds 90%, GPT-Live-1 starts a replacement voice engine within the same session seeded with original instructions + up to 8,192 tokens of history (recent messages + summary). session.instructions.append / session.thinking.append / session.commentary.append inject content with acknowledged start_ms/end_ms windows on the session timeline.
• Transcripts: session.input_transcript.delta and session.output_transcript.delta deliver timed fragments for UI, logging, moderation and early-start work; the API also reports cumulative session.usage.updated (~once per minute; final usage at session.closed).
• Audio formats: WebSocket PCM16 mono 24 kHz (audio/pcm) or G.711 μ-law 8 kHz (audio/pcmu) for direct telephony bridging (e.g., Twilio Media Streams); WebRTC negotiates its own format. Session durations billed per second and not rounded up; a WebRTC creation bills 15s of initialization that is credited once the session starts (cost-optimization guide).
• Data controls: GPT-Live-1 is eligible for Zero Data Retention (with limitations; stored sessions ignored under ZDR), processes in US or EU/EEA+Switzerland regions (2026 guide data per Progressive Robot), and session recording (store:true) is off by default, retained 30 days when enabled.

• Real-time voice is now a commodity API building block: $0.05/min per-second billing for ChatGPT-Voice-grade full-duplex conversation makes "an AI that can hold a phone conversation" a routine engineering cost, not a research project — the same commoditization pattern as text tokens in 2023–24.
• The voice/brain split is an architectural statement: GPT-Live-1 deliberately separates conversation from reasoning. OpenAI is telling developers to stop treating voice as a model feature and start treating it as a layer you can swap while keeping your tools, guardrails and state machine. This is a different design center than the Realtime API (one model does all) and a direct challenge to incumbent speech stacks (Twilio-style orchestration with STT/LLM/TTS chains).
• It lands inside the week's live-voice platform race: same window as Google's Gemini 3.8 Live (Sep 15) and xAI's Voice Agent API at $0.08/min (Sep 17, Galaxy Day 3). OpenAI's $0.05/min voice layer sets the price reference point; Google answers with assistant-default distribution, xAI undercuts on telephony-agent positioning.
• Pairing with the Agents API (same day) and Presence completes the enterprise story: OpenAI now sells (a) a managed enterprise voice agent (Presence), (b) the raw voice model (GPT-Live-1 API), and (c) a managed reasoning harness (Agents API beta) — a vertical stack for customer-facing voice automation that competes with telecom/contact-center incumbents (Twilio, Five9, NICE) even as Twilio integrates it.
• It corrects the record on "multimodal voice": the GPT-Live-1 API accepts no images and no video (audio+text only). Anyone building "see the world" agents must continue using the Realtime API (image input) or other multimodal models — a meaningful constraint for the week's "live-video agent" narrative that the discovery record incorrectly attributed to GPT-Live-1.

• Full-duplex phone agents in a weekend: reservations, order status, customer support and outbound calls where the caller can interrupt, change direction and talk over the agent — previously a multi-vendor integration project (STT+VAD+LLM+TTS+telephony), now a session.start call plus a Twilio Media Stream bridge (Twilio shipped the native GPTLiveProvider the same day).
• Bring-your-own-brain voice: pair GPT-Live-1 with Luna for cheap high-volume scheduling, Terra for balanced work, Astra for hard reasoning, or a third-party/self-hosted model via client delegation — voice quality is decoupled from reasoning capability, so teams can swap backends without touching the voice layer.
• Voice-enabled coding and ambient agents: the launch post demonstrates GPT-Live-1 + Codex SDK (talk to a repo-reading agent), and the cost guide sketches ambient agents that close the voice session while a long task runs and resume on completion ("Resume conversation" button).
• Natural-language telephony UX: barge-in handling, "mhmm" backchannels, silence tolerance, and keyword-biased alphanumeric capture (order IDs, dates) — the interaction grammar of real phone calls rather than IVR menus.
• Language/voice localization: 12 accent/dialect/language voices at launch with more promised, letting products localize voice agents as a configuration change rather than a TTS vendor swap.

• Will GPT-Live-1 mini (and Medium/High reasoning variants) ship in the API, at what price, and when? (DevDay Sep 29 is the near-term candidate.)
• Will the Realtime API line get a deprecation date now that GPT-Live-1 is GA — forcing migration for gpt-realtime-2.1 users?
• Can user-side speech transcripts actually be retrieved via the API (discrepancy between model-card claims and community testing)? OpenAI's clarification would settle a production blocker.
• When will video and screen-sharing arrive for GPT-Live (OpenAI said "soon" at the July launch; still absent Sep 10)? Its arrival would finally justify the discovery record's "video" framing — retroactively wrong, but directionally predictive.
• Will OpenAI publish independent third-party full-duplex benchmarks (or open Full Duplex Bench/Tau3 methodology) to back the +30pp and #1 claims?
• How will voice-layer pricing evolve once Google (Gemini 3.8 Live) and xAI ($0.08/min) publish comparable benchmark numbers?
• Does OpenAI Presence become the enterprise distribution layer that sidelines self-serve GPT-Live-1 API deployments, or do both scale?

• Near term (days–weeks): developer community pressure on OpenAI to (a) expose user-side transcripts, (b) add session.cancel-like controls, (c) confirm Realtime deprecation timing; expect documentation updates and possible API tweaks before OpenAI DevDay on Sep 29 (Fort Mason, San Francisco) — the natural venue for GPT-Live-1 mini, video support announcements, and benchmark methodology releases.
• This week's competitive clock: Google's Gemini 3.8 Live (Sep 15) and xAI's Voice Agent API (Sep 17) give buyers three full-duplex voice surfaces within 7 days; benchmark shootouts (Artificial Analysis-style) are imminent.
• Data/legal follow-ups: OpenAI published a GPT-Live system card (with a corrected/re-run safety evaluation noted Aug 4); voice-agent deployments will start showing up in EU AI Act GPAI transparency reporting and GDPR audio-processing audits.
• Product trajectory: expect voice-layer price compression, a mini-tier for low-cost telephony, video/screen-sharing parity with Advanced Voice Mode ("working to introduce these capabilities soon" per July launch materials), deeper Codex/Agents-API integration, and Foundry/enterprise adoption as the GPT-6 Astra GA (Sep 17, S06) normalizes frontier models on Azure.

Report this story as "OpenAI opens its full-duplex real-time voice model to all developers — GPT-Live-1 hits the API at $0.05/minute" — NOT as a video model. The correction matters: GPT-Live-1 accepts audio and text only, and OpenAI is explicit that image and video are unsupported, so the "real-time video agent" angle belongs to other models (Realtime with image input, or future GPT-Live updates). The genuinely newsworthy substance is (1) ChatGPT-Voice-grade full-duplex conversation as a commodity API at a legible $0.05/min per-second price; (2) the architectural bet that voice and reasoning should be separable layers — a reframing of how voice agents are built, with real migration costs for Realtime users; (3) the same-week collision with Google's Gemini 3.8 Live and xAI's $0.08/min Voice Agent API, making this the week real-time voice became a platform war; and (4) the pairing with the Agents API beta and OpenAI Presence, completing OpenAI's enterprise voice stack. Flag the benchmark claims as vendor-measured, and surface the early community-reported production gaps (transcript retrieval, hangup reliability) honestly — they are the difference between a demo and a phone line.