Most voice assistants take turns: they wait for you to stop, think, then speak. A full-duplex model listens and speaks at the same time, so it can acknowledge you mid-sentence and stop the moment you interrupt. On October 1, 2026, Tavus previewed Griffin-Lite, a full-duplex model for video calls. Five models now make that claim in their own documentation, and they differ sharply in access, licence and what their makers have measured.
Editorial note: Built from each maker’s own announcement, model card or repository, read October 3, 2026. Every latency and benchmark figure here is the maker’s own; none was reproduced by Digital Applied.
- 01Five models qualifyOpenAI’s GPT-Live-1, Tavus’s Griffin-Lite and three open-weight models from NVIDIA and Kyutai.
- 02Only one is a paid APIGPT-Live-1 costs $0.05 a minute for the voice layer. Griffin-Lite is a closed preview.
- 03Latency claims do not line upEach maker measures something different, on different hardware. Test on your own calls.
- 04Open models are runnableVoiceChat, PersonaPlex and Moshi can run on your own GPUs, with licences to check first.
01 — DefinitionWhat full-duplex means
The term comes from telephony: a full-duplex line carries sound both ways at once. A traditional voice agent, as OpenAI describes it, chains three systems: speech recognition, a language model, then speech synthesis. That chain has to decide when you have finished before it can start, which is how pauses get cut off and interruptions land late.
A full-duplex model handles both audio streams in one model. Kyutai’s Moshi repository describes the design plainly: one stream for the model’s own speech, one for the user’s, modelled together. OpenAI makes the business case in its GPT-Live-1 announcement: one model reasoning over incoming and outgoing audio avoids the delays and fragile handoffs of a chained system. It cites the language-learning company Speak, which found interruptions fell by almost 80% against earlier turn-based systems, a customer-reported figure.
02 — The newsWhat changed this week
Tavus announced Griffin on October 1 as a model that sees, hears and talks at once in a video call, and released Griffin-Lite as a research preview to a small group of testers. It decides what to do at sub-second intervals rather than once per turn, and generates the video face as well as the voice. Tavus reports that 26 of 54 people (48%) in its own study thought they had spoken to a real person after a one-minute call; that is Tavus’s study, with Tavus’s method.
On September 30, Inworld announced it had acquired Ultravox, a platform for building real-time voice agents. Inworld says the deal advances its work on speech-to-speech experiences and that more will follow in the coming weeks. Its announcement does not describe Ultravox as full-duplex, so Ultravox is not in the tables below.
03 — The dataThe models that do it
Each row is a model whose maker states that it listens and speaks at the same time. Two are commercial; three publish their weights.
| Model | Access | Price or licence | What it adds |
|---|---|---|---|
| GPT-Live-1 (OpenAI) | Closed API; first introduced in ChatGPT | $0.05 a minute for the voice layer, plus the backend model | Delegates reasoning and tool calls to a separate model; phone calls supported |
| Griffin-Lite (Tavus) | Research preview for selected early testers | No public price | Video as well as voice: sees the user and generates a talking face |
| NemotronLabs VoiceChat 11B (NVIDIA) | Open weights on Hugging Face; NVIDIA container for live use | OpenMDW 1.1 licence | Tool calls during the conversation; English |
| PersonaPlex 7B (NVIDIA) | Gated open weights; accept the licence on Hugging Face | Weights under NVIDIA’s Open Model License; code MIT | Set the role with a text prompt and the voice with an audio prompt |
| Moshi (Kyutai) | Open weights; code and setup on GitHub | Weights CC-BY 4.0; code MIT and Apache | The open base PersonaPlex is built on |
The two NVIDIA models answer different needs. The VoiceChat model card says it is the first open full-duplex model to call tools, and lets a developer set a holding phrase the agent speaks while a tool runs. The PersonaPlex repository focuses on persona: a text prompt sets the role, such as a customer service agent, and an audio prompt sets the voice. GPT-Live-1 takes a third route and hands hard questions and tool calls to a separate text model, which our GPT-Live-1 API guide covers in detail.
04 — The dataThe latency each maker reports
Latency is the number everyone quotes and nobody measures the same way. The table gives each figure with what it measures and where. Read across a row, never down a column.
| Model | Reported figure | What and where |
|---|---|---|
| GPT-Live-1 | 30 percentage points better on Full Duplex Bench than GPT-Realtime-2.1 | OpenAI’s own evaluation; no millisecond figure in the announcement text |
| Griffin-Lite | 0.43 s from audio arriving to its effect on screen | Tavus, measured on H100s, for the video generator |
| NemotronLabs VoiceChat | 448 ms turn-taking; 480 ms to yield when interrupted | NVIDIA, Full-Duplex-Bench 1.0, on H100 |
| Moshi | 160 ms theoretical; as low as 200 ms in practice | Kyutai, on an L4 GPU |
| PersonaPlex | No figure in the public README | NVIDIA points to FullDuplexBench prompts for evaluation |
Moshi’s figure is its overall latency on an L4 GPU. NVIDIA’s is time to reply after a user finishes, on a data-centre GPU. Tavus’s covers video, not speech. OpenAI gives a benchmark gain, not a time. Putting them in one chart would rank four different measurements. Our voice agent latency reference separates the delays worth timing.
05 — MethodHow to test one yourself
The public benchmark most of these makers cite, Full-Duplex-Bench, scores four behaviours, and they make a good test plan for your own calls. Record 20 real conversations from your use case and score each model on the same four.
Pause handling
Pause mid-thought for two seconds. A good model waits; a bad one answers half a question.
Backchannels
Talk for 30 seconds. Listen for short acknowledgements that do not take the turn.
Turn-taking
Time from the end of your sentence to its first word, on your network and hardware.
Interruption
Cut in mid-answer. Time how long it keeps talking, and check it kept the context.
Add one test the benchmark does not cover: background noise. OpenAI lists noise handling as a strength of GPT-Live-1, and it is where phone-line deployments fail first. Benchmark rankings and listener preference also disagree, as our review of voice benchmarks against listener votes found, so include a few real listeners in the scoring.
06 — Practical implicationsWhich one fits
Full duplex fixes the rhythm of a call, not its substance. The agent still needs the right tools, data and limits behind it, and a plan for when it is wrong. Our AI transformation work covers voice agents from model choice to call-flow testing.
Score two models on 20 of your own calls
Pick one paid and one open model from the tables, run the four tests plus background noise on 20 recorded conversations, and let a few real listeners rate the results. That answers more than any number a maker has published.