GPT-Live : OpenAI launches full-duplex voice — AI agents can finally listen and speak at the same time

GPT-Live : OpenAI launches full-duplex voice — AI agents can finally listen and speak at the same time

🔎 Voice was just a gadget, it becomes a controller

For years, voice interaction with AI hit an invisible ceiling. You speak, the AI waits for you to stop talking, it processes, then it responds. This half-duplex operation — borrowed from walkie-talkies — turned every voice exchange into a succession of siloed monologues. Even ChatGPT's advanced voice mode, despite being a real breakthrough, suffered from this structural latency: interruption remained a hack, not a feature.

On July 9, 2026, OpenAI announced GPT-Live, a family of voice models designed natively in full-duplex. The AI listens and speaks simultaneously, like a human. You can interrupt it, bounce off ideas, clarify in real time while it is delivering its response. No waiting latency, no "please, go on".

The timing is not coincidental. GPT-Live arrives on the heels of GPT-5.6 Sol, ChatGPT Work, and the rise of autonomous AI agents. OpenAI is not just launching a new voice mode. It is positioning voice as the primary control surface for a new generation of agents — agents that run in the background, as shown by the Ona acquisition, and that can be piloted by voice without ever touching a keyboard.

The essentials

• GPT-Live is a family of natively full-duplex voice models: the AI listens and speaks at the same time, without artificial turn-taking.
• Two versions launched: GPT-Live-1 (maximum performance) and GPT-Live-1 mini (reduced latency, optimized cost), both accessible via API.
• The model breaks the half-duplex paradigm that limited Siri, Alexa, and even ChatGPT's previous voice mode.
• OpenAI is explicitly targeting voice control of AI agents: hands-free supervision, continuous instructions, fluid multimodal interaction.
• Competition is active: ElevenLabs crosses the 500 million dollar ARR mark, Google's Gemini Live already offers full-duplex, but GPT-Live's agent integration differentiates the offering.

Recommended tools

Half-duplex vs full-duplex: it's not just a technical detail

The difference between half-duplex and full-duplex is the fundamental divide between a tech demo and an interface usable in everyday life.

In half-duplex, the communication channel is unidirectional at any given instant. One speaks, the other listens. When you use Siri or ChatGPT's standard voice mode, the system detects the end of your sentence (vad — voice activity detection), sends the audio to the model, waits for generation, and then plays the response. This process introduces an incompressible latency of 1 to 3 seconds between your last syllable and the AI's first word. Worse: if you interrupt, the system must cancel its current generation, reset the context, and start over. The interruption is not natural; it is an error case handled with varying degrees of grace.

In native full-duplex, two audio streams flow simultaneously in both directions. The AI continuously generates its spoken response while continuing to receive and process your incoming audio stream. If you say "no, rather..." in the middle of its sentence, the model adjusts its trajectory in real time — exactly like a human who corrects their statement when their listener nods or contradicts them.

This is an architectural change, not an optimization. Half-duplex stacks patches (interruption detected → stop → re-prompt) onto a system designed for sequential processing. Full-duplex rethinks the model itself to handle two parallel streams as its normal state.

The OpenAI o1 System Card paper had already documented the challenges related to managing intermediate states in reasoning models. GPT-Live extends this issue to the audio channel: the model must maintain a coherent reasoning state while continuously integrating new auditory stimuli. This is fundamentally more complex than a text model, because real time is unforgiving — every additional millisecond of latency destroys the illusion of natural conversation.

GPT-Live-1 and GPT-Live-1 mini: two models, two strategies

OpenAI is not offering a single model but a range structured around two different compromise points.

GPT-Live-1 is the flagship model. It is optimized for scenarios where the complexity of voice reasoning is critical: negotiation, consulting, diagnostics, supervision of complex agents. Its latency is in the range of 200-300 ms, which remains perceptible but is well below the breaking point of a natural conversation (generally placed around 500 ms by computational linguistics studies).

GPT-Live-1 mini sacrifices some reasoning depth to gain in reactivity. Its target latency drops below 150 ms, making it suitable for quick voice commands, mainstream mobile interfaces, and scenarios where speed takes precedence over nuance. Its inference cost is also significantly lower, making it viable for large-scale integrations.

This duality reflects a clear market strategy: mini for volume, the standard model for value. This is the same logic that structures the range of best LLMs for coding, where lightweight models coexist with heavy reasoning models depending on the use case.

Access is exclusively via API. No dedicated ChatGPT interface at launch, which signals that OpenAI is initially targeting developers and integrators, not the direct general public. Monetization is through usage, not through an additional subscription.

Why This Is a Leap for AI Agent UX

Full-duplex voice isn't about making conversations more friendly with a chatbot. It's about controlling systems that work in the background.

Let's take the concrete scenario of an agent executing long tasks — precisely the type of workflow that ChatGPT Work enables. You ask the agent to research information, compile a report, contact providers. In text mode, you have to alternately check the interface and go back to your keyboard. In half-duplex voice mode, you spend your time saying "ok continue" and waiting.

With GPT-Live in full-duplex, you can say "start with the first three, and if the last one doesn't respond within two hours, move on to the next" while the agent is listing its plan of action. You don't leave your flow. You don't wait for your turn to speak. Voice becomes a continuous supervision channel, like a project manager giving directives while walking through the office.

It's this agent + voice integration that changes the game. Voice alone is a gadget. Voice as an agent controller is an interface.

The implications for interaction design are profound. Developers must rethink their patterns: no more "press the button, wait for the response," but continuous flows where the user can modulate, correct, redirect in real time. The best autonomous AI agents that currently operate in headless mode gain a natural interface that requires no screen.

The strategic context: price war and agent war

GPT-Live is not arriving in a vacuum. It is part of an offensive sequence from OpenAI that began in early July 2026.

First, GPT-5.6 Sol with its Terra and Luna variants, which marked the start of an aggressive price war. Then the acquisition of Ona (formerly Gitpod), which transformed Codex into a system of persistent agents capable of running even when your laptop is closed. Then ChatGPT Work, the agent that executes tasks for hours. And now GPT-Live, which adds the voice layer.

The logic is coherent: OpenAI is building a complete vertical stack. The base model (GPT-5.6), the execution infrastructure (Ona/Codex), the agent framework (ChatGPT Work), and the control interface (GPT-Live). Each layer reinforces the others. Voice without the agent remains a chatbot. The agent without voice remains confined to the text interface.

This strategy puts enormous pressure on the competition. Google has Gemini Live, which already offers full-duplex and integrates with the Android ecosystem. But Google does not yet have the equivalent of ChatGPT Work in terms of long-duration persistent agents. ElevenLabs dominates voice synthesis and has crossed the 500 million ARR mark, but remains a voice provider, not an agent provider. The voice + agent convergence is the space OpenAI is looking to occupy.

Concrete use cases: beyond the demonstration

Automated customer service

Customer service is the most obvious use case, but also the most demanding. Current voice chatbots are universally hated because they are slow, rigid, and unable to handle the unexpected. Half-duplex condemns them to a linear script: "say 1 for..., say 2 for...".

In full-duplex, a voice agent based on GPT-Live can handle a call like a human. The customer can interrupt to clarify their problem, the agent adjusts in real time, does not repeat its entire script, and can seamlessly move on to a resolution without reconnecting. The difference in experience is comparable to going from a paper form to an actual phone conversation.

The challenge is no longer technical but organizational: companies must integrate these agents into their backend systems (CRM, tickets, knowledge bases) so that vocal fluidity does not stop at the model's border.

Continuous personal assistant

The voice personal assistant has always failed on the same hurdle: latency and the inability to handle continuous context. You ask something, the assistant answers, and everything starts from scratch.

GPT-Live, combined with agent persistence via the Ona infrastructure, enables an assistant that stays present. You can talk to it while walking, interrupt to answer a colleague, and resume the conversation without repeating the context. The assistant maintains its reasoning state in the background, just as a meilleur LLM pour les agents IA would do in text, but with the voice channel as the interface.

Voice control of development tools

Less visible but potentially more transformative: the voice control of development tools. A developer coding with an agent based on Codex can give continuous voice instructions while the agent generates code. "No, not that library, use the other one" — said while the agent is writing the import. Full-duplex makes this possible without friction.

This is particularly relevant for the meilleurs LLM pour coder like GPT-5.3 Codex, which scores 80 on the agentic benchmark. Voice becomes a natural complement to the keyboard, not a replacement.

Competition: ElevenLabs, Gemini Live and the rest of the market

The voice AI market in July 2026 is far from being an OpenAI monopoly.

ElevenLabs remains the undisputed leader in raw vocal quality. Their voice cloning, emotion handling, and library of synthetic voices have no equivalent. But ElevenLabs is a component provider (synthesis, recognition), not an end-to-end conversational model provider. Their strength lies in the sound, not in the reasoning. Integration with reasoning models like GPT-Live or the best LLMs on the market is done through assembly, which adds latency.

Google's Gemini Live is the most direct competitor on paper. Google already offers full-duplex voice interaction integrated into Android and the Google ecosystem. Google's advantage is device-native integration: no need for APIs, no integration complexity. The disadvantage is the closed ecosystem and the lack of an agent framework as mature as what OpenAI is building with ChatGPT Work.

Open source models are also starting to explore this direction. For developers who want to keep total control, running voice models locally with Ollama or via open source AI agents with Ollama is becoming a credible alternative, even if full-duplex capabilities still lag behind proprietary models.

The comparison table is clear:

Implications for developers

For developers, GPT-Live changes the rules of voice integration.

Until now, building a fluid voice experience required stacking three distinct components: an automatic speech recognition (ASR) model, a language model for reasoning, and a text-to-speech (TTS) model. Each component added its own latency, and coordinating between the three was a complex engineering exercise. Interruption, in particular, required hacks: detecting the user's speech, canceling the ongoing TTS, reinjecting the context into the LLM, and restarting generation.

GPT-Live unifies these three layers into a single model. ASR, reasoning, and TTS are no longer separate bricks but integrated capabilities of the same system. For the developer, this means a single API, a conversation state managed by the model, and interruption that is a native behavior, not an exception to handle.

Cost remains a point of attention. Audio models are significantly more expensive to infer than text models. GPT-Live-1 mini exists precisely to address this constraint: for high-volume, low-complexity use cases, the mini is the economical choice. For high-value-added scenarios (advisory, diagnostics, negotiation), the standard model justifies its extra cost.

Developers who are already using the meilleurs LLM gratuits to prototype will need to integrate GPT-Live into their production stack. The good news: the API is designed to replace an existing ASR+LLM+TTS pipeline without rearchitecting the application.

❌ Common mistakes

Mistake 1: Confusing full-duplex with "fast interruption"

Many comments following the announcement reduced GPT-Live to "a voice mode where you can interrupt faster." This is inaccurate. Fast interruption in half-duplex remains an interruption: the model stops, resets, and re-generates. In full-duplex, there is no interruption in the strict sense — there is continuous modulation of a stream that never stops. The distinction is technical, but its consequences on UX are radical.

Mistake 2: Thinking that voice replaces text

Full-duplex voice is complementary to text, not a substitute. For complex queries involving tabular data, code, or formal precision, text remains superior. GPT-Live excels in supervision, real-time modulation, and hands-free use cases. The mistake is trying to force voice into scenarios where text is more suitable.

Mistake 3: Ignoring network latency cost

Model latency (150-300 ms) is only part of the problem. Network latency between the user and the OpenAI API often adds an extra 50-200 ms depending on location. For consumer applications, neglecting this network component leads to disappointing experiences in real-world conditions, even if the model is technically capable of full-duplex.

Mistake 4: Underestimating audio bias management

Voice models inherit the biases of their training data. Accents, speech rates, language registers — GPT-Live is not exempt from these issues. Ignoring the diversity of input voices leads to agents that work perfectly for a standard American accent and degrade for everything else. Testing must include a variety of voice profiles.

❓ Frequently Asked Questions

Does GPT-Live replace ChatGPT's current voice mode?

Not immediately. At launch, GPT-Live is only accessible via API. Integration into the ChatGPT interface is likely but not yet announced. The current voice mode remains in place for general consumers.

What is the difference with Google's Gemini Live?

Both offer full-duplex, but GPT-Live is designed to integrate natively with the OpenAI agent ecosystem (ChatGPT Work, Codex). Gemini Live is more mature in its Android mobile integration, but less connected to a persistent agent framework.

Does GPT-Live work in French?

Yes, the model supports French among other languages. However, as with any multilingual voice model, quality and latency may vary compared to English. The meilleurs LLM en français are sometimes still more performant in text than in voice for the language of Molière.

Can GPT-Live be used locally?

No. GPT-Live is a proprietary model only accessible via the OpenAI API. For local alternatives, you need to turn to ASR + LLM + TTS combinations run locally, but native full-duplex is not yet available in the open-source ecosystem at this level of quality.

Which reasoning model does GPT-Live use in the backend?

OpenAI has not published details of the internal architecture. It is likely that GPT-Live relies on variants of the GPT-5.x family (GPT-5.4, GPT-5.5) adapted for continuous audio processing, but this remains speculative in the absence of detailed technical documentation.

✅ Conclusion

GPT-Live is not a voice update — it is the missing control interface for the generation of autonomous agents that OpenAI has been building since early 2026. Native full-duplex transforms voice from a gadget into a continuous supervision channel, and the timing with ChatGPT Work and the acquisition of Ona leaves little doubt about the strategy. If the quality holds up in production, the question will no longer be "why use voice with an AI agent" but "why go without it". To follow the evolution of these models, check out our monthly comparison of the best LLMs.