OpenAI pushed its GPT-Live-1 voice API into general developer availability on Wednesday, transplanting the full-duplex conversational engine behind ChatGPT Voice onto programmable infrastructure. Priced at $0.05 per minute for the audio layer, with backend reasoning billed separately, the release puts the same architecture that serves roughly 150 million weekly ChatGPT Voice users into the hands of any team building a voice product. The move sharpens competition with Google’s Gemini Live API, ElevenLabs, Vapi, Retell, Hume, and the broader field of voice-agent providers now forced to differentiate on a direct architectural comparison rather than incremental latency wins.
GPT-Live-1 voice API: Why Full-Duplex Changes the Voice Stack
For most of the past decade, voice agents have run on cascaded pipelines. A speech-to-text transcriber listens for silence, hands a transcript to a language model, and the model’s text reply is converted back to audio by a separate synthesizer. Each hop adds latency, and each handoff sheds the tone, emotion, and prosody that color human speech. Worse, the silence detector cannot reliably tell a thoughtful pause from background noise, which is why early voice bots interrupted speakers mid-sentence. Published P50 latency on these stacks runs from about 800 milliseconds to two seconds in clean conditions, and real-world telephony deployments routinely stretch further once carrier round-trip overhead is included.
The GPT-Live-1 voice API collapses that three-stage chain into a single model that processes incoming and outgoing audio simultaneously. The system makes speaking-versus-listening decisions multiple times per second without waiting for a silence gap, eliminating the voice-activity-detection trigger that defines cascaded architectures. OpenAI also trimmed session startup from six network round trips to one, a meaningful gain for telephony where call-connection time directly shapes caller satisfaction.
Delegation, Benchmarks, and Cost Posture
Live conversation and deep reasoning pull in opposite directions, and the architecture addresses that with a delegation layer. The fast conversational surface stays on GPT-Live-1, while expensive tasks like web search, repository-aware queries, or multi-step tool execution are routed to a backend model that runs asynchronously and folds its results back into the live session. Developers pick the backend, which can be Codex, GPT-6 Astra, a third-party model, or an internal lightweight model for high-volume commodity calls like scheduling. The cost arithmetic improves: frontier-compute charges apply only to tasks that need them, not every conversational beat.
On Full Duplex Bench, which measures pause handling, turn-taking, interruptions, and backchanneling, GPT-Live-1 outperformed the prior GPT-Realtime-2.1 by roughly 30 points, with the largest gains in turn-taking latency and interactive behavior. On Tau3, an end-to-end spoken-task benchmark covering airline, retail, telecom, and banking support flows, the model paired with GPT-6 Astra at medium reasoning effort scored 86.2 percent, against 45.7 percent for GPT-Realtime-2.1 and 42.4 percent for GPT-Realtime-2.0. Both scores derive from OpenAI’s own evaluation rather than independent audit, and Tau3 spans scripted task domains rather than the full noise, accent, and dialect diversity of production telephony.
Early Partners Cite Concrete Engineering Wins
OpenAI released the API alongside a slate of early-partner numbers that frame the upgrade in operational terms. Speak, the language-learning platform, tracked false interruptions, where a voice system misreads a learner’s thinking pause as the end of a turn, and reported an 80 percent reduction against prior turn-based systems. Co-founder and CTO Andrew Hsu compared the result to the judgment required of a skilled tutor: restraint matters as much as intervention. Yelp’s reservation assistant, Yelp Host AI, showed lifts in call-handling rates, and CTO Alex Levy pointed to a behavioral shift in callers, who now speak in fuller, more natural sentences, a qualitative signal that the experience no longer feels like navigating an IVR tree.
An unnamed healthcare company offered the most operationally dramatic figure. The team reported an 80 percent codebase reduction after retiring the glue code that stitched transcriber, reasoner, and synthesizer together, retiring roughly 23,000 lines of orchestration logic. Co-founder and CTO Tony Stoyanov framed the savings as engineering capacity redirected away from plumbing toward the patient experience itself. Intercom’s Fin support agent and Cognition’s Devin coding assistant also lit up as early adopters, with Intercom COO Jordan Neil describing the move from stop-start cadence toward the natural rhythm of a phone call, and Cognition co-founder Walden Yan characterizing Devin voice sessions as a teammate-like collaboration that survives a hands-off, away-from-keyboard context.
The Voice Stack Decision Developers Face Now
Voice teams now face a specific architectural decision rather than a vague direction-of-travel question. Cascaded stacks from Vapi at $0.05 per minute plus model costs, ElevenLabs at $0.08 per minute, Retell at $0.07 to $0.31 per minute by configuration, and Hume’s Empathic Voice Interface at $0.05 to $0.07 per minute remain viable options, particularly for teams with existing transcription or telephony investments. The new variable is whether full-duplex capability, sub-second turn latency, and the delegation cost pattern justify the migration. Switching costs accrue quickly once a platform decision is made, because caller expectations, agent tooling, and analytics pipelines all adapt to the chosen interaction model, and replacing that substrate later carries the same weight as a database migration. For developers evaluating the GPT-Live-1 voice API, the practical checklist now includes benchmark provenance, the strength of delegation hooks into existing backend models, telephony startup latency under real carrier conditions, and the willingness of callers to attribute the smoother pacing to improved artificial intelligence rather than an interruption rate. Early partner data points to meaningful gains on each axis, though independent production deployments across accent and noise diversity will determine whether the benchmark lead translates into a sustained platform advantage.
Source: Full-Duplex Voice AI Hits Developer APIs: GPT-Live-1 Kills Cascaded Pipeline