GPT-Live-1 full-duplex voice AI OpenAI released GPT-Live-1 to developers on Wednesday, putting the full-duplex voice model behind ChatGPT Voice into its developer API at $0.05 per minute for the voice layer, with backend reasoning and agent infrastructure billed separately. The release moves full-duplex voice from a consumer-facing feature into selectable developer infrastructure, putting the architecture that reaches 150 million weekly ChatGPT Voice users in direct competition with Google’s Gemini Live API, ElevenLabs, Vapi, Retell AI, and Hume AI. Developers can now invoke the same single-model full-duplex design that powers ChatGPT Voice, forcing every voice-agent team to weigh integration costs against an architectural switch that compounds once made.
GPT-Live-1 full-duplex voice AI replaces cascaded pipelines
Traditional voice agents stitch three discrete models together: a speech-to-text transcriber waits for silence, hands a transcript to a language model, then routes text to a speech synthesizer. Each handoff adds latency; each transition strips tone and emotional cues before the model reasons about the conversation. Because the stack waits for a silence gap to trigger each stage, it cannot tell a thinking pause from ambient noise, which produced the false interruptions that plagued earlier AI voice products.
Published P50 latency for cascaded systems runs from roughly 800 milliseconds to two seconds under optimal conditions, the source notes, and real-world telephone deployments add 80 to 150 milliseconds of carrier round-trip overhead. P99 latency can be orders of magnitude worse under load. GPT-Live-1 collapses that pipeline into a single unified model that processes incoming and outgoing audio simultaneously, the same full-duplex pattern the telecommunications industry has used since the earliest telephone networks, distinct from half-duplex walkie-talkie behavior 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, eliminating the turn-trigger latency of voice-activity detection entirely. OpenAI’s infrastructure team also cut voice-session startup from six network round trips to one, a meaningful change for telephony deployments where call initiation speed affects customer experience.
GPT-Live-1 full-duplex voice AI uses a delegation layer for reasoning
Full-duplex conversation and deep reasoning pull in opposite directions: conversation requires millisecond-scale responses, while complex web search, tool execution, or multi-step reasoning cannot finish in milliseconds. GPT-Live-1 resolves that tension with a delegation layer. The model manages the live conversational surface, making fast decisions about what to say next, when to stay quiet, and when to acknowledge, while routing heavier tasks asynchronously to a backend model. The backend runs while GPT-Live-1 keeps talking, and the result folds back into the conversation when ready.
Developers pick 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 through a session.commentary.append call. The same pattern supports GPT-6 Astra, a third-party model, or a lightweight in-house model for high-volume commodity tasks like scheduling or order updates. The cost implication is that teams can match reasoning depth and price to the task instead of paying for frontier-level computation on every conversational exchange.
GPT-Live-1 full-duplex voice AI 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 gains in turn-taking latency and interactive behavior. On Tau3, which measures voice-agent intelligence across end-to-end spoken tasks including 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%. That compared to 45.7% for GPT-Realtime-2.1 and 42.4% for GPT-Realtime-2, placing it ranked first on Tau3 overall.
The source flags that these evaluations come from OpenAI rather than independent auditors, and that Tau3 measures success on scripted task domains. Real-world performance across the noise, accent, and query diversity of production telephone traffic will depend on deployments at scale.
GPT-Live-1 full-duplex voice AI early partner results
OpenAI published early-partner results alongside the API. Speak, the language-learning platform, focused on false interruptions, the moments when a voice system cuts a learner off 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 framed the result in tutoring terms, saying that 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 that callers now speak in fuller, more natural sentences, an indicator that the conversational experience feels different from a conventional IVR or turn-based voice bot. An unnamed healthcare company provided the most operationally concrete figure: switching to GPT-Live-1 simplified its codebase by 80 percent and eliminated 23,000 lines of glue code previously needed to stitch together speech recognition, reasoning, and synthesis in a cascaded pipeline. Co-founder and CTO Tony Stoyanov said the reduction freed engineers to focus on the patient experience rather than the plumbing. Intercom’s Fin support agent and Cognition’s Devin coding agent also joined as early adopters. Intercom COO Jordan Neil described the result as moving voice support from a stop-start rhythm toward the natural flow of a phone call, while Cognition co-founder and CPO Walden Yan said working with Devin via GPT-Live-1 begins to feel like collaborating with a teammate who can talk through ideas, pressure-test approaches, and hand off work while the user is away from the keyboard.
GPT-Live-1 full-duplex voice AI pricing and competitive landscape
At $0.05 per minute for the voice layer alone, GPT-Live-1 is price-competitive with Vapi, which charges $0.05 per minute plus separate model costs on a cascaded pipeline. 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 prices its Empathic Voice Interface at $0.05 to $0.07 per minute depending on tier. Google’s Gemini Live API offers full-duplex architecture through the Gemini API at no cost on its free developer tier, with usage limits scaling across paid tiers, supports telephony through FreeSWITCH integration, and adds multimodal inputs including video and screen sharing that GPT-Live-1 does not currently offer. NVIDIA released VoiceChat-11B, an open-weights model with roughly 450-millisecond turn-taking latency, aimed at self-hosted deployments.
The architectural divide is the most consequential comparison. Full-duplex models such as GPT-Live-1, Gemini Live, NVIDIA VoiceChat-11B, and Sesame process audio simultaneously in both directions and decide in context when to respond. Cascaded platforms including ElevenLabs, Vapi, and Retell chain separate speech-recognition, language-model, and speech-synthesis services with voice-activity detection as the turn trigger. Cascaded stacks offer modularity, letting developers swap components, but carry irreducible handoff latency and remain fragile under real-world noise and hesitation patterns. OpenAI also expanded from a small voice set to twelve voices, including Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, and Cinder, with custom voice access through the sales team. Telephony support is native, and in late July 2026 OpenAI added SynthID audio watermarking to GPT-Live output, embedding an invisible, transformation-resistant watermark developers can verify through the Content Provenance Check API. GPT-Live-1 full-duplex voice AI.
What to watch
GPT-Live-1 turns full-duplex voice into infrastructure that any developer team can invoke by the minute, shifting the competitive benchmark for voice agents from demo naturalness to real-world call quality and developer experience at scale. The immediate questions for the next several months are whether the 30-point Full Duplex Bench gain and 86.2% Tau3 score hold across the noise, accent, and call-duration distribution of production telephony, how Gemini Live’s free tier and multimodal inputs affect share among cost-sensitive builders, and whether OpenAI Presence, the enterprise voice platform built on GPT-Live-1, becomes the default deployment path for larger customers. Audio provenance through SynthID watermarks will also matter as enterprises start tracking AI-generated audio in customer interactions, keeping GPT-Live-1 full-duplex voice AI at the center of both the architectural and the governance conversation.