OpenAI Launches GPT-Live-1 API for Full-Duplex Voice Apps That Talk and Listen Simultaneously

OpenAI has opened GPT-Live-1, its full-duplex speech model, to developers via API. The model can listen and speak at the same time rather than waiting for a user to finish talking before responding—a capability OpenAI calls full-duplex interaction. GPT-Live-1 already powers voice mode inside ChatGPT and is now available for third-party integration.

Pricing is set at $0.05 per minute, which OpenAI acknowledges is not a low cost for high-volume voice applications. Developers can pair GPT-Live-1 with different backend reasoning models depending on the task, allowing them to trade off reasoning depth, response speed, and cost per use case.

Benchmark Gains Over GPT-Realtime-2.1

According to OpenAI's own benchmarks, GPT-Live-1 shows substantial improvements over its predecessor, GPT-Realtime-2.1:

• Full-duplex interactivity score: 80.1% vs. 45.4%
• Turn-taking latency: 0.8 seconds vs. 1.4 seconds
• Tool-calling accuracy: 87% vs. 60%
• Banking voice support pass rate: 32% vs. 12.4%

These figures come from OpenAI's internal testing and have not been independently verified. The banking voice support benchmark in particular—despite the improvement—still shows a relatively low absolute pass rate of 32%, suggesting real-world reliability in complex, regulated voice interactions remains limited.

Twelve New Voices, Built-In Transcripts

GPT-Live-1 ships with twelve new voices covering a range of accents, dialects, and languages. The API also outputs ASR (automatic speech recognition) transcripts alongside response text by default, giving developers text logs of both sides of a conversation without additional processing. OpenAI says full technical details will be published in its API documentation.

Early Deployment: Yelp

Yelp is already using GPT-Live-1 for phone-based reservation handling. According to Yelp CTO Alex Levy, the model has improved call handling quality, though no specific metrics were disclosed to quantify the improvement.

What This Means

Full-duplex capability—responding while still listening, rather than in strict turn-based exchanges—addresses one of the most persistent weaknesses in voice AI: unnatural pauses and awkward interruptions that make automated calls feel robotic. A drop in turn-taking latency from 1.4 seconds to 0.8 seconds is significant for phone-based use cases like reservations, customer support, and appliance-style voice agents, where delay directly affects user experience.

The jump in tool-calling accuracy, from 60% to 87%, is arguably more consequential for developers than the latency figures. Voice agents that need to check inventory, pull account data, or execute transactions depend on reliable function calls; a 27-point improvement narrows the gap between demo-quality voice bots and production-ready ones.

At $0.05 per minute, however, cost will be the limiting factor for scale deployments. A contact center handling thousands of daily calls will face meaningful per-minute expenses that require automation gains to offset. The relatively modest 32% pass rate on the banking benchmark also signals that even with improvements, GPT-Live-1 is not yet reliable enough for high-stakes, regulated voice interactions without human fallback.