GPT-Live-1 API Is Live: Full-Duplex Voice for Developers

OpenAI launched GPT-Live-1 in its API on September 10, giving developers the full-duplex voice model that has powered ChatGPT Voice since 2025 at $0.05 per minute for the voice layer. OpenAI reported the model improves full-duplex interactivity to 80.1% from 45.4% on GPT-Realtime-2.1, cuts turn-taking latency to 0.8 seconds from 1.4 seconds, and raises tool-calling accuracy to 87% from 60%. Yelp deployed GPT-Live-1 through its Yelp Host product after handling more than 1 million calls since October 2025, and one healthcare company said switching removed 23,000 lines of code, 80% of its voice agent implementation.

OpenAI launched GPT-Live-1 in its API on September 10, giving developers the same full-duplex voice model that has powered ChatGPT Voice since 2025. The model listens and speaks simultaneously — no chained speech-to-text pipeline, no hand-off delays, no custom interruption logic. At $0.05 per minute for the voice layer, the headline price is accessible. The performance jump over GPT-Realtime-2.1 is real: 80.1% full-duplex interactivity versus 45.4%, turn-taking latency down to 0.8 seconds from 1.4 seconds, and tool-calling accuracy up to 87% from 60%. How the GPT-Live-1 API Works Traditional voice agents are really three fragile models holding hands: speech-to-text, then language model, then text-to-speech. Every handoff introduces latency and a chance to lose the conversational rhythm. GPT-Live-1 replaces that chain with a delegation pattern. The voice model handles the acoustic layer — listening, speaking, interruption detection, turn-taking — while complex reasoning is passed to a backend model of the developer’s choice. OpenAI surfaces this through a dedicated endpoint v1/live/sessions with WebRTC, WebSocket, and SIP telephony transport options. The model supports function calling mid-conversation, so a restaurant reservation system can look up availability and confirm a booking without ever dropping the call. Developers choose the backend: GPT-6 Astra for complex reasoning, Luna for high-volume budget tasks, or any model they prefer. A healthcare company put a number on what simplification looks like: switching to GPT-Live-1 removed 23,000 lines of code from their codebase — 80% of their voice agent implementation. That deleted code was pipeline orchestration, custom interruption logic, and synchronization glue. GPT-Live-1 handles all of it natively. As The New Stack noted https://thenewstack.io/gpt-live-1-voice-api/ , the 80% figure is specific to that deployment, but the engineering complexity reduction is real and measurable. The Numbers Worth Caring About Benchmarks only matter if they correlate with real outcomes. Yelp deployed GPT-Live-1 through its Yelp Host product after handling more than 1 million calls https://ppc.land/yelp-and-hatch-put-openais-gpt-live-1-into-voice-ai-after-1-million-calls/ since October 2025. Their CTO’s observation is more telling than any benchmark: callers are now speaking in “fuller, more natural sentences.” That is a signal the model no longer cuts off mid-thought or sounds robotic enough to change how people talk. The numbers from OpenAI’s official announcement https://openai.com/index/introducing-gpt-live-1-in-the-api/ back this up. Full Duplex Bench improved by 30 points. Tau3 Voice Intelligence tasks hit 86.2% task success. Banking support pass rate reached 32% versus 12.4% for GPT-Realtime-2.1. The latency difference is what matters most in practice: 0.8-second turn-taking feels conversational. 1.4 seconds does not. That gap is why GPT-Live-1 is a meaningful upgrade, not a marginal one. Related: OpenAI Agents API Public Beta: Build Without the Boilerplate https://byteiota.com/openai-agents-api-public-beta-build-without-the-boilerplate/ GPT-Live-1 API Pricing: What It Actually Costs $0.05 per minute covers the voice layer only. Every backend model call, tool invocation, and reasoning step is billed separately at standard Responses API rates. For a voice agent that does real work — confirming an order status, querying a booking database, or running a multi-step customer verification — the backend costs will routinely exceed the voice layer charge. OpenAI prices sessions per second rather than per full minute, which helps for short calls. However, developers need to model both components before assuming GPT-Live-1 is cheap at scale. A 5-minute call with frequent tool calls can cost substantially more than $0.25. The voice layer price is a floor, not the ceiling. Teams migrating from existing Realtime API implementations should expect their cost model to change. GPT-Live-1 API: What Developers Need to Know Before Building GPT-Live-1 is not a drop-in replacement for gpt-realtime-2.1. It uses a different endpoint v1/live/sessions , not the Realtime or Chat Completions endpoints, and the integration path is distinct. Developers evaluating both should check the official API documentation https://developers.openai.com/api/docs/models/gpt-live-1 before assuming existing Realtime code migrates cleanly. Three constraints matter at launch: the model’s knowledge cutoff is July 31, 2025, so any real-time information requires tool calls; concurrent session limits are 25 on Tier 1 and 500 on Tier 5, making Tier 1 unsuitable for production traffic; and the free tier is not supported. Plan tier upgrades before launch, not after. The SIP telephony transport is a genuine advantage for teams replacing traditional IVR systems — direct phone integration without additional telephony middleware. Key Takeaways - GPT-Live-1 replaces the STT-LLM-TTS pipeline with a single full-duplex model — the 23,000-line code reduction at one deployment is a real measure of the engineering simplification - The performance gap over GPT-Realtime-2.1 is substantial: better latency, higher tool accuracy, and Yelp’s 1M+ calls confirm production viability - $0.05 per minute covers only the voice layer — backend model calls and tool usage are separate charges that will often cost more than the voice fee itself - Plan for concurrency limits 25 on Tier 1 , the July 2025 knowledge cutoff, and the distinct API endpoint before starting a migration or new build