What is GPT-Live and how it differs from everything before

GPT-Live is built on a full-duplex architecture. This means the model doesn't operate in discrete turns ("you speak, I respond"), but continuously — many times per second — evaluates whether to speak, listen, pause, interrupt you, or call a tool. During the conversation, it naturally accompanies you with short sounds like "mhmm" or "right," signaling that it's paying attention to what you're saying.

This is the third developmental stage of OpenAI's voice models. The first generation — cascaded systems (the original ChatGPT Voice) — chained three separate models: speech-to-text transcription, a language model, and text-to-speech synthesis. Information was lost between models, response was slow and unnatural. The second generation — turn-based models (Advanced Voice Mode) — processed audio in a single model, but still waited for you to finish. A short pause or background noise often caused the model to interrupt you mid-thought.

GPT-Live removes these limitations with two architectural changes:

1. Continuous interaction (full-duplex) — the model processes input and generates output simultaneously. It decides many times per second what to do next. It can let you think without interrupting. It also supports live translation.
2. Delegation of deeper work — when you need to search the web, reason logically, or perform a more complex task, GPT-Live hands it off to GPT-5.5 in the background. While GPT-5.5 is working, GPT-Live continues conversing with you naturally. Once the result is ready, it returns it to the conversation. This architecture also means GPT-Live will automatically leverage future OpenAI models as well.

GPT-Live-1 vs GPT-Live-1 mini: Two variants for different needs

OpenAI is launching two versions. GPT-Live-1 is the default model for ChatGPT Go, Plus, and Pro subscribers. GPT-Live-1 mini is a lightweight variant for free-tier users. Both models begin rolling out globally today on iOS, Android, and the web at ChatGPT.com.

Interestingly, users can choose a reasoning level:

• Instant — fast responses, uses GPT-5.5 Instant in the background
• Medium — balanced performance, uses GPT-5.5 Thinking with medium reasoning level
• High — maximum answer quality, uses GPT-5.5 Thinking with high reasoning level

How GPT-Live performs in tests

OpenAI conducted extensive comparative tests, measuring GPT-Live-1 and GPT-Live-1 mini against Advanced Voice Mode. In head-to-head tests where evaluators compared 5–10 minute conversations, both new models were significantly preferred across all metrics: overall impression, turn-taking, interruptions, fluency, and naturalness of interaction.

On automated benchmarks, GPT-Live-1 also outperforms Advanced Voice Mode:

• GPQA — expert-level scientific reasoning (biology, chemistry, physics): significant improvement
• BrowseComp — agentic web search and finding hard-to-locate information: strong increase
• τ³-Voice Telecom — realistic multi-turn customer support tasks: GPT-Live-1 wins

What changes in ChatGPT Voice for regular users

Over 150 million people weekly use ChatGPT Voice and dictation. For them, several tangible improvements launch today:

• More natural conversations — you can interrupt, pause, or ask ChatGPT to slow down. The model naturally responds with words like "mhmm" or "I see." OpenAI also remastered all nine voices for GPT-Live.
• Smarter responses — voice mode now uses OpenAI's latest frontier models, including GPT-5.5.
• Better listening — when you pause to think, ChatGPT waits. When you ask it to just listen, it does. In noisy environments (a passing car, surrounding conversations), it focuses better on your voice.
• Visual cards — during a call, ChatGPT can display cards with weather, stocks, sports scores, and other information. Voice mode continues to support search, memory, images, and file uploads.

What's still missing

At launch, GPT-Live does not support video or screen sharing in ChatGPT — OpenAI plans to deliver these features later. Users who need these features can continue using older versions of ChatGPT Voice (Standard and Advanced Voice Mode).

The API for developers is also not yet available, but OpenAI is already collecting registrations from interested parties and plans to make it available soon. For some languages, the model may have a non-native accent or gaps in fluency — OpenAI is actively working on improvements.

The competition isn't sleeping: Google Gemini, Anthropic, and others

OpenAI isn't alone in the voice AI field. Google recently launched Gemini 3.1 Flash TTS with high expressiveness and voice controllability, and at I/O 2026 introduced Gemini Omni with video editing. Anthropic, meanwhile, is investing in Claude's voice capabilities. Competitive pressure is enormous — and that's exactly why OpenAI is coming with an architecturally bold solution that doesn't just rely on a bigger model, but on smarter design.

For Czech users, the key question is language support. GPT-Live is optimized for "the most popular languages in ChatGPT," but OpenAI admits that for some less represented languages, the accent may be non-native. Czech is among the officially supported languages for ChatGPT Voice, but its quality in full-duplex mode remains to be seen in practice.

Safety: Real-time guarding

GPT-Live underwent extensive safety testing including audio-native evaluations, synthetic tests on risk areas (self-harm, psychosis, emotional dependency on AI, violence, sexual content), and red-teaming by internal experts. Compared to Advanced Voice Mode, it performed comparably or better across virtually all evaluated areas.

The model has built-in safeguards that can intervene even during speech — redirecting the conversation to a safer response, displaying a warning, or in extreme cases ending the call. For teenage users, there are parental controls, and the model was directly trained on age-appropriate behavior.