What happened

OpenAI just opened up an AI voice model that can talk and listen at the same time to outside developers. The model is called GPT-Live-1, and as of September 10, 2026, anyone building an app can plug into it through an API (that's just a set of instructions that lets one piece of software talk to another).

The headline feature is "full-duplex" audio. Think about a regular phone call: normally, one person talks while the other stays quiet, then they swap. Full-duplex means the AI can listen while it's still talking, the same way two people do in a real conversation, where you can jump in, interrupt, or say "mm-hmm" without derailing things. GPT-Live-1 is already running inside ChatGPT's voice mode, and now the same engine is available for anyone to build with.

Developers can also mix and match: GPT-Live-1 handles the actual speaking and listening, while a separate "reasoning" model does the thinking behind the scenes. That means a company can pick a cheaper, faster brain for simple tasks and a smarter, slower one for complicated requests, all through the same voice interface.

On OpenAI's own benchmark tests, GPT-Live-1 beats its predecessor, GPT-Realtime-2.1, by a wide margin. In tests measuring how well it handles full-duplex conversation, it scores 80.1 percent versus 45.4 percent for the older model. The gap it leaves before responding, called turn-taking latency, drops from 1.4 seconds to 0.8 seconds. That difference is what separates a conversation that flows from one that feels like a walkie-talkie.

Tool-calling accuracy, meaning how reliably the AI can actually take an action while talking, like pulling up an account balance or booking a slot, jumps from 60 percent to 87 percent. In a banking customer-support test, GPT-Live-1 passed 32 percent of the time, up from 12.4 percent for the previous model. That's a real improvement, but it also shows the technology still fails on most complex calls.

The model ships with twelve new voices covering different accents, dialects, and languages, plus built-in transcripts of both what the caller said (ASR, or automatic speech recognition, is the tech that turns spoken words into text) and what the AI replied. Pricing is $0.05 per minute, which OpenAI itself flags as not cheap. Yelp is already using it for phone-based restaurant reservations; the company's CTO, Alex Levy, says it's handling calls better than before.

What it means for you

At home: voice assistants that don't cut you off mid-sentence. Ask your smart speaker a question, then change your mind halfway through, and it can adjust instead of waiting for you to finish and start over.

At work: phone-based customer support that sounds and behaves more like a real person, including the ability to be interrupted, corrected, or asked a follow-up without breaking the flow.

Running a business: small businesses without a big call center, restaurants, salons, clinics, can hand off appointment booking and reservations to an AI phone line, the same way Yelp is testing it for restaurant bookings.

Studying: language-learning apps can offer a conversation partner that responds instantly and in different accents, which is closer to talking with a real native speaker than the stiff, turn-based bots most apps use now.

Creative work: podcasters, video creators, and audiobook producers get access to a wider range of natural-sounding voices for narration, dubbing, or interactive audio experiences.

MyKreaTool AI chat — try ChatGPT, Claude and Gemini in one place. Free on MyKreaTool.Open the tool →

Extra income: freelancers and small agencies can build and sell voice-agent setups to local businesses, restaurants, dentists, real estate offices, that want an AI receptionist but don't have an engineering team.

How to try it right now

You don't need to be a developer to get a feel for AI voice tools first.

1. Start free. Head to mykreatool.com and try its free AI tools to experiment with AI-generated voice and speech before spending anything. It's the fastest way to understand what this kind of technology sounds like and where it still stumbles.

2. If you're building something real, go to OpenAI's developer platform and request access to the GPT-Live-1 API.

3. Read the API documentation OpenAI publishes for GPT-Live-1 before writing any code, so you know exactly which backend models it pairs with and what data it returns.

4. Test with a small batch of calls or interactions first. Since pricing is $0.05 per minute, run a short pilot before committing to full-scale use.

5. Watch the transcripts. GPT-Live-1 returns both the speech-to-text transcript and its own response text automatically, so you can review exactly what happened on every call.

Upsides and what changes

The biggest shift is that conversations with AI stop feeling like a walkie-talkie exchange and start feeling like an actual back-and-forth. Faster turn-taking (0.8 seconds instead of 1.4) and much higher tool-calling accuracy (87 percent versus 60 percent) mean AI phone agents can both sound natural and actually get things done, like checking an order status or rescheduling an appointment, mid-conversation. Twelve voices across different accents and languages also make it easier to match an AI voice to a specific audience instead of using one generic-sounding assistant everywhere.

Limitations

Be honest about where this still falls short: at $0.05 per minute, cost adds up fast for any business running high call volumes, and the 32 percent pass rate on the banking support benchmark means GPT-Live-1 still fails on most complex, high-stakes conversations, so it's not ready to fully replace trained human staff for anything sensitive like financial disputes. It also requires real developer work to integrate through the API; there's no plug-and-play consumer app version yet outside of ChatGPT's own voice mode.

Conclusion

GPT-Live-1 makes AI conversations feel less like a walkie-talkie and more like talking to a person, but it's priced and built for developers, not casual users, and it still fails most complex calls. One action for today: try a free AI voice tool to hear the natural, full-duplex difference for yourself before deciding whether a paid, developer-built solution is worth it for your business.