OpenAI released GPT-Live-1 to developers on Wednesday, moving the full-duplex voice model that powers ChatGPT Voice into its developer API and triggering direct competition with Google’s Gemini Live API, ElevenLabs, Vapi, and every other voice-agent stack in the market. Any developer can now build a voice product on the same architecture that reaches 150 million weekly ChatGPT Voice users, at $0.05 per minute for the voice layer. The backend reasoning model and agent harness are billed separately.

The release matters because it transforms full-duplex voice from a consumer feature into choosable developer infrastructure — the kind of platform decision that, once made, accumulates switching costs fast. Teams now face a concrete architectural question: adopt GPT-Live-1’s single-model approach, or continue building on the cascaded pipelines most voice agents still use today.

How Three-Stage Pipelines Break — and What GPT-Live-1 Does Instead

To understand what changed, it helps to understand what the old architecture actually does. Traditional voice agents string three discrete models together: a speech-to-text transcriber listens until it detects silence, then hands a text transcript to a language model, which generates a text reply, which a text-to-speech synthesizer speaks aloud. Every step adds latency. Every handoff loses information — tone and emotional cues disappear before the model ever reasons about the conversation. And because the system waits for silence to trigger each stage, it cannot distinguish a genuine pause from a user thinking mid-sentence from ambient noise, leading to the unwanted interruptions that plagued earlier AI voice products.

Published P50 latency for cascaded systems runs from roughly 800 milliseconds to two seconds under optimal conditions; real-world telephone deployments, which add 80 to 150 milliseconds of carrier round-trip overhead, push that higher — and P99 latency can be orders of magnitude worse under load.

GPT-Live-1 collapses that architecture into a single unified model. It processes incoming and outgoing audio simultaneously — what the telecommunications industry called full-duplex since the earliest telephone networks, as distinct from the half-duplex walkie-talkie model where only one direction transmits at a time. The model decides many times per second whether to speak, pause, listen, interject a brief acknowledgment, or fire a tool call, without waiting for a silence gap to trigger the next stage. That eliminates the turn-trigger latency of voice activity detection entirely.

OpenAI’s infrastructure team reduced voice-session startup from six network round trips to one. That number matters for telephony deployments, where call initiation speed directly affects customer experience.

Delegation Architecture: How Reasoning Fits Into a Real-Time Conversation

Full-duplex conversation and deep reasoning pull in opposite directions: conversation requires millisecond-scale responses; complex web search, tool execution, or multi-step reasoning cannot happen in milliseconds. GPT-Live-1 resolves this with a delegation layer. The model handles the live conversational surface — the fast decisions about what to say next, when to stay quiet, when to acknowledge — while routing tasks that need more processing to a backend model running asynchronously. The backend works while GPT-Live-1 keeps the conversation going; the result folds back into the conversation when ready.

Developers choose the backend model. A code example published with the release shows how to connect GPT-Live-1 to Codex — passing conversation context, running a repo-aware query, and returning the answer mid-session via a session.commentary.append call. The same pattern works with GPT-6 Astra, a third-party model, or a lightweight in-house model for high-volume commodity tasks like scheduling or order updates. The practical cost implication: teams can match reasoning depth and price to the actual task rather than paying for frontier-level computation on every conversational exchange.

Benchmark Results

On Full Duplex Bench — which evaluates pause handling, turn-taking, interruptions, and backchanneling — GPT-Live-1 beat GPT-Realtime-2.1 by 30 points, with the largest measured gains in turn-taking latency and interactive behavior.

On Tau3, which measures voice-agent intelligence across end-to-end spoken tasks — customer-service scenarios in airline, retail, and telecom domains, plus spoken banking-support tasks requiring knowledge retrieval and account-management tools — GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort scored 86.2%, compared to 45.7% for GPT-Realtime-2.1 and 42.4% for GPT-Realtime-2. That placed it ranked first on Tau3 overall.

A note on benchmarks: these evaluations come from OpenAI, not from independent auditors. Tau3 measures success on specific scripted task domains; real-world performance across the full noise, accent, and query diversity of production telephone traffic will depend on deployments at scale.

What Developers Are Already Reporting

OpenAI released the API alongside a set of early-partner results that illustrate both what the model does well and where the gains are concrete enough to quantify.

Speak, the language-learning platform, built one of the more specific analyses. Its engineering team focused on false interruptions — the moments when a voice system cuts a learner off mid-thought because it mistakes a thinking pause for the end of a turn. In early evaluations, GPT-Live-1 reduced false interruptions by 80 percent compared with previous turn-based systems. Co-founder and CTO Andrew Hsu described the result in terms of what it takes to be a skilled human tutor: knowing when to hold back is as important as knowing when to step in.

Yelp deployed the model through its Yelp Host AI reservation system. CTO Alex Levy reported improvements in call-handling rates and noted a behavioral signal from callers: people are now speaking in fuller, more natural sentences — an indicator that the conversational experience on the other end feels genuinely different from a conventional interactive voice response or turn-based voice bot.

An unnamed healthcare company provided perhaps the most operationally concrete figure: switching to GPT-Live-1 simplified their codebase by 80% and eliminated 23,000 lines of code previously needed to stitch together real-time patient conversations — the glue code that wires together speech recognition, reasoning, and synthesis in a cascaded pipeline. Co-founder and CTO Tony Stoyanov said the reduction freed engineers to work on the patient experience itself rather than the plumbing.

Intercom’s Fin support agent and Cognition’s Devin coding agent also joined as early adopters. Jordan Neil, COO of Intercom, described the result as moving voice support “from the stop-start rhythm toward the natural flow of a phone call.” Walden Yan, co-founder and CPO of Cognition, said working with Devin via GPT-Live-1 begins to feel like collaborating with a teammate — talking through ideas, pressure-testing approaches, handing off work while away from the keyboard.

Voice AI Stack Decision: What Developers Are Actually Choosing Between

GPT-Live-1’s arrival opens a specific platform comparison that any developer building a voice product now has to make.

At $0.05 per minute for the voice layer alone, GPT-Live-1 is price-competitive with Vapi, the developer-focused telephony platform that uses a cascaded pipeline and charges $0.05 per minute plus separate model costs. ElevenLabs, known for voice cloning quality, uses a cascaded turn-taking architecture at $0.08 per minute. Retell AI, which targets call-center automation, runs a cascaded stack at $0.07 to $0.31 per minute depending on configuration. Hume AI, which emphasizes emotional awareness in speech-to-speech models, prices its Empathic Voice Interface at $0.05 to $0.07 per minute depending on tier. As reviewed in independent voice agent cost comparisons, total cost across all these platforms varies significantly by how backend model usage is bundled above the base voice-layer price.

Google’s Gemini Live API offers full-duplex architecture via the Gemini API and is available at no cost on its free developer tier, with usage limits scaling across paid tiers. It supports telephony deployments through Gemini Live API supports telephony via FreeSWITCH integration and multimodal inputs including video and screen — a capability GPT-Live-1 does not currently offer. NVIDIA released VoiceChat-11B open-weights model with roughly 450-millisecond turn-taking latency, targeting self-hosted deployments.

The technical distinction that matters most in this comparison is architectural. Full-duplex models process audio simultaneously in both directions and make in-context decisions about when to respond — this is how GPT-Live-1, Gemini Live, NVIDIA VoiceChat-11B, and Sesame operate. Cascaded platforms such as ElevenLabs, Vapi, and Retell chain separate speech-recognition, language-model, and speech-synthesis services, with voice-activity detection as the turn trigger. The cascaded approach offers more modularity — developers swap out any component — but carries irreducible latency from the handoffs and remains fragile under real-world noise and hesitation patterns. A developer who chooses today will accumulate integration costs that make switching non-trivial.

Twelve Voices, Telephony Support, and Audio Provenance

For the API release, OpenAI expanded from a small set of voices to twelve across accents, dialects, and languages: Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, and Cinder. Custom voice access requires contacting OpenAI’s sales team.

Telephony support is native — GPT-Live-1 can be deployed for phone-based voice agents handling inbound calls from restaurant reservations to customer support lines, without additional infrastructure for the full-duplex layer.

In late July 2026, OpenAI added SynthID audio watermarking to GPT-Live output. This embeds an invisible, transformation-resistant watermark into the audio waveform of AI-generated voice. Developers and organizations can use OpenAI’s Content Provenance Check API to verify whether a given audio clip was generated by an OpenAI model — a capability relevant for teams that need to track AI-generated audio in customer interactions or documentation.

For enterprises that want a more integrated deployment path, OpenAI is directing customers toward OpenAI Presence enterprise voice platform, which uses GPT-Live-1 to power real-time agents that can answer questions, resolve issues, operate within company systems, take approved actions, and escalate to human agents as needed.

Does This Change the Voice AI Market?

GPT-Live-1’s API release represents a structural shift: full-duplex voice is no longer a consumer-facing demo — it is infrastructure that any developer team can invoke by the minute. The competitive benchmark for voice agents moved from “does it sound natural in a demo” to “which API platform delivers the best developer experience and real-world call quality at scale.”

Whether GPT-Live-1 sustains its benchmark advantage in production — across a realistic distribution of noise conditions, accents, call durations, and query types — will take months of developer deployments to validate. Partners operating in the first weeks of access describe meaningful improvements in specific, measurable outcomes. Whether those gains hold across every vertical and telephony environment where voice agents are deployed is a question the market will answer, not OpenAI’s benchmarks.

Frequently Asked Questions

What is full-duplex voice AI and why does it matter for developers?

Full-duplex voice AI can listen and generate speech simultaneously, the same way a telephone call works — both parties can speak or hear at any moment. Older voice AI systems used a half-duplex, turn-based model: the system waited for the user to stop speaking (detected by a voice-activity detection algorithm), processed the audio, and replied. That architecture introduced meaningful latency at every stage and was brittle when users paused to think or when background noise triggered a false end-of-turn. For developers, full-duplex means they can build voice agents that handle interruptions, deliver real-time acknowledgments, and feel conversationally natural — without manually coordinating between separate speech-recognition, language-model, and text-to-speech services. The telecommunications industry called full-duplex the standard for telephone communication since its earliest days, and GPT-Live-1 brings that model to developer AI infrastructure.

How much does GPT-Live-1 cost, and what does the price include?

GPT-Live-1 is priced at $0.05 per minute for the front-end voice layer — the model handling listening, speaking, and real-time decision-making. Backend reasoning is billed separately based on which model a developer chooses: pairing with GPT-6 Astra costs more than pairing with a lighter, faster model suited to simpler tasks. Total cost depends on how much of the conversation requires deep reasoning versus routine handling. At $0.05 per minute for voice layer pricing and breakdown, GPT-Live-1 is price-competitive with cascaded alternatives like Vapi that also price the voice layer at $0.05 per minute and bill backend costs separately, while ElevenLabs starts at $0.08 per minute and Retell AI runs from $0.07 to $0.31 per minute depending on configuration.

Can GPT-Live-1 be used for phone calls, and how does it compare to Gemini Live?

Yes — GPT-Live-1 includes telephony support for inbound and outbound phone-based voice agents, handling restaurant reservations, customer support lines, and similar use cases natively. Google’s Gemini Live API is the most direct full-duplex competitor and is available at no cost on its developer free tier, making it attractive for teams that want to evaluate full-duplex architecture without upfront commitment. Gemini Live also offers multimodal inputs (video, screen sharing) that GPT-Live-1 does not currently support. The meaningful comparison for production workload call quality comparison is not just price or features but real-world call quality across noise, accent diversity, and long sessions — which neither platform’s published benchmarks fully capture.

Does GPT-Live-1 mark AI-generated audio so listeners can identify it?

OpenAI added Google DeepMind’s SynthID audio watermark into GPT-Live output in late July 2026. SynthID embeds an invisible watermark directly into the audio waveform — a signal designed to survive common audio transformations like format conversion or compression. Developers can use OpenAI’s Content Provenance Check API to verify whether a given audio clip carries an OpenAI provenance signal. The watermark does not affect audio quality and is not an audible identifier; it is a technical provenance marker for platforms and tools that integrate with C2PA content-credential standards.