OpenAI's GPT-Live talks and listens simultaneously

Key Takeaways

• GPT-Live uses full-duplex architecture to listen and speak simultaneously, eliminating the rigid turn-taking of previous voice modes
• Complex queries get delegated to GPT-5.5 in the background, boosting accuracy from 0.7% to 75.2% on agent-based web search benchmarks
• Users preferred GPT-Live-1 over Advanced Voice Mode in 75.7% of cases during OpenAI's evaluations

OpenAI has released GPT-Live, a voice model that can listen and speak at the same time. The full-duplex architecture eliminates the awkward pauses of previous ChatGPT voice modes, where users had to wait for the AI to finish before responding. Two versions are rolling out globally: GPT-Live-1 for paying subscribers on Go, Plus, and Pro plans, and GPT-Live-1 mini for free accounts.

The bigger shift is architectural. When a question requires web search, reasoning, or agent capabilities, GPT-Live hands the task to GPT-5.5 running in the background. The conversation keeps flowing while the heavier model works. OpenAI says the system is built so GPT-Live stays connected to whatever frontier model is current.

How does full-duplex change AI voice interaction?

Traditional voice assistants, including ChatGPT's older Advanced Voice Mode, use half-duplex audio. The AI waits for you to stop talking, processes your input, then responds. Real human conversation doesn't work that way. People overlap, interrupt, and give verbal nods like "mhmm" to signal they're following along.

GPT-Live makes decisions multiple times per second about whether to speak, listen, pause, or interrupt. It uses filler phrases to acknowledge the speaker without taking over the conversation. Users can cut in mid-sentence or ask the model to slow down.

Nvidia released a similar open-source model called PersonaPlex earlier this year, but OpenAI's implementation ships directly to ChatGPT's 100 million+ weekly users across iOS, Android, and the web.

The benchmark gap between GPT-Live and Advanced Voice Mode

The performance difference is stark. On GPQA, a scientific reasoning benchmark, GPT-Live-1 at high reasoning level hits 84.2% accuracy. Advanced Voice Mode scores 45.3%. The gap widens on BrowseComp, which tests agent-based web search. GPT-Live-1 reaches 75.2%. Advanced Voice Mode manages 0.7%.

That 0.7% figure explains why previous voice modes felt limited. They answered questions using only their own parameters, with no ability to search the web or delegate to more capable models. For anything beyond simple queries, they were unusable for real work.

OpenAI's internal tau3 Voice Telecom test measures full-duplex voice agents on realistic telecom support tasks. GPT-Live-1 at high reasoning completes about 65% of tasks in roughly 385 seconds. Advanced Voice Mode hits 30% and takes just as long. The instant reasoning tier finishes 37% of tasks in around 225 seconds, trading accuracy for speed.

User preference and the reasoning level trade-off

In OpenAI's evaluations, users preferred GPT-Live-1 over the previous Advanced Voice Mode in 75.7% of comparisons. The mini version, available on free accounts, was preferred 69.2% of the time.

Users can now select a reasoning level. "Instant" delivers quick answers with less background processing. "Medium" and "High" tell ChatGPT to spend more time thinking, routing more work to GPT-5.5. The trade-off is latency versus accuracy, and users get to choose based on the task.

What's missing at launch

GPT-Live doesn't yet support video or screen sharing. The older Standard and Advanced Voice Modes, which include those features, will remain available for now. OpenAI says video and screen sharing are coming soon.

API access isn't live yet either. Developers can sign up through a form, and OpenAI says it plans to add API support soon. Until then, builders who want to integrate GPT-Live into their own products will have to wait.

The update also adds visual cards during voice conversations. ChatGPT can now display weather forecasts, stock prices, or sports scores on screen while talking. OpenAI revamped the nine available voices for GPT-Live as well.

The safety question around human-sounding AI

More natural AI conversation raises safety concerns. Research shows people tend to treat AI models like humans and can be talked into actions with serious consequences. Heavy voice model users appear even more susceptible to this effect.

OpenAI hasn't detailed specific safety measures for GPT-Live, but the pattern is clear: as voice AI gets more convincing, the line between tool and trusted advisor blurs. Product teams building on this technology will need to think carefully about how users perceive their AI's suggestions.

Frequently Asked Questions

What is GPT-Live?

GPT-Live is OpenAI's new voice model for ChatGPT that uses full-duplex architecture to listen and speak simultaneously, making conversations feel more natural than previous voice modes.

Is GPT-Live free to use?

GPT-Live-1 mini is available on free ChatGPT accounts. GPT-Live-1, the more capable version, requires a Go, Plus, or Pro subscription.

Can GPT-Live search the web during conversations?

Yes. GPT-Live delegates web searches and complex reasoning tasks to GPT-5.5 in the background while maintaining the conversation.

When will GPT-Live API access be available?

OpenAI says API access is coming soon. Developers can sign up through a form on OpenAI's website to join the waitlist.

Does GPT-Live support video calls?

Not at launch. Video and screen sharing features are planned but not yet available. The older Advanced Voice Mode still supports these features.

Source: The Decoder / Matthias Bastian