August 4, 2026. OpenAI published an engineering post on August 3 explaining how it built GPT-Live, its third generation voice system, in six months. The post is unusually specific for a vendor write up, and the specificity is the story: almost none of the responsiveness win came from the model. It came from deleting a component, rewriting the transport, and redefining what capacity means.
What OpenAI disclosed
- The turn detector is gone from the audio path. Earlier voice systems used a small model to guess when the user had stopped speaking. Guess early and you cut people off, guess late and the reply drags. GPT-Live's voice model is full duplex, listening and speaking at the same time, which OpenAI says removes the need for a separate detector.
- Thinking is delegated, not inlined. When deeper reasoning or a tool call is required, GPT-Live consults a frontier model such as GPT-5.5 on a separate asynchronous path. A slow tool call delays its own result rather than stalling the conversation.
- The media frontend was rewritten in Go. Replacing a Python asyncio implementation improved frame delivery enough that, in OpenAI's own phrasing, "the new system's p95 matching the previous system's p50".
- Session startup fell from six network round trips to one. OpenAI published WARP, the WebRTC Abridged Roundtrip Protocol, as an open specification, alongside a technique it calls Instant Connect that takes the signalling exchange off the critical path. Together they let a client start a session with a single UDP packet. WARP support is already in libwebrtc and Pion, and the drafts are being advanced through the IETF.
- Capacity was redefined after a silent production test. Shadow routing real traffic showed a supporting component saturating before the GPU did. OpenAI changed its planning question from how many requests a GPU can handle to "how many concurrent sessions can the system sustain while keeping every frame on schedule".
What it means for operators
If you are buying or specifying a voice agent for sales calls, reception or qualification, this post doubles as a free evaluation checklist. Ask the vendor three questions. Does the agent rely on a turn detector? What happens when a caller interrupts mid sentence? Does a slow CRM or calendar lookup pause the conversation while it runs? Those separate systems that sound alive from systems that sound like a phone tree, and none of them are answered by naming a model. Geography counts too, since OpenAI reports that routing a session to distant capacity adds delay at several points, which matters if your callers sit outside the region your provider serves.
The caveat is timing
OpenAI describes the GPT-Live API as upcoming rather than available, so there is nothing to build against yet. If you need a voice agent live this quarter, ship on what exists today, whether that is your GoHighLevel stack or a dedicated voice provider, and hold the next vendor to this architecture as the standard. When we scope voice into an AI automation build, interruption handling and end to end latency get tested before anything else, because they are the first things a caller notices.
Frequently Asked Questions
GPT-Live is OpenAI's third generation voice system. Its voice model is full duplex, meaning it listens and speaks at the same time, and it delegates deeper reasoning and tool use to a frontier model on a separate asynchronous path. OpenAI says it powers ChatGPT Voice and that a GPT-Live API is upcoming.
A turn detector is a small model that guesses when a speaker has finished. Guessing too early cuts the caller off and guessing too late makes the reply feel sluggish, and the larger model could not start work until the detector decided. Taking it out of the audio path removes that trade off, which is why interruption handling is a fair question to ask any voice vendor.
Not directly. OpenAI's August 3, 2026 engineering post describes the GPT-Live API as upcoming rather than generally available, so any production voice agent you need this quarter has to run on something that already ships.
Whether the system uses a turn detector, what happens when a caller talks over the agent, whether a slow lookup pauses the conversation, and where the inference capacity serving your callers is located. Those four answers predict call quality better than the name of the underlying model.