Editorial illustration for OpenAI's GPT-Live-1 API Enables Simultaneous Talk-and-Listen Apps
OpenAI's GPT-Live-1 API Enables Real-Time Talk
OpenAI opened up GPT-Live-1 to developers this week as an API, giving outside teams access to the same speech model already running inside ChatGPT. The headline feature is full-duplex audio: the model can talk and listen at once, rather than waiting for a pause before responding, which is how most voice bots still work. Developers can mix and match backend models depending on what a task needs, balancing reasoning power, response speed, and cost.
That flexibility comes at a price. OpenAI is charging $0.05 per minute for GPT-Live-1, and the company is already pointing to a real customer to justify the cost. Yelp has started using the model to handle phone-based reservations, and CTO Alex Levy says it's improved how the company manages those calls.
The bigger question is how much GPT-Live-1 actually improves on what came before it, particularly GPT-Realtime-2.1, OpenAI's previous voice model. OpenAI has published its own benchmark numbers comparing the two across interactivity, latency, and task accuracy, along with results from a banking voice support test meant to simulate real customer service conditions.
OpenAI is making GPT-Live-1 available to developers as an API. The speech model can listen and talk at the same time, a feature known as "full-duplex," and is already running inside ChatGPT.
Why this matters
For developers building voice apps, the pricing tells you as much as the benchmarks do. Five cents a minute is a real cost line item, not a rounding error, and it forces a build decision that wasn't necessary before: do you need full-duplex, or were you fine with turn-based models that cost less? Yelp's use case, phone reservations, makes sense for the premium.
A bot that can listen while it talks avoids the dead air and interruption problems that make automated calls feel obviously automated. That's worth paying for if you're handling thousands of calls a day.
The bigger signal is the architecture choice. Letting developers pair GPT-Live-1 with different backend models means OpenAI is pushing people toward tiered systems, cheap and fast for simple exchanges, heavier reasoning models for anything complicated. We'd watch how many teams actually build that way versus defaulting to one model for everything.
The benchmark jump from 45.4 to 80.1 percent on full-duplex interactivity is the kind of number that looks great in a slide deck. Whether it holds up on a noisy customer service line is the real test.
Related Reading
Browse all OpenAI news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source
Related Topic OpenAI Calls Apple's Lawsuit 'Rotten to Its Core' in Rebuttal
Policy & Regulation
From Archives Full-Stack NIM Optimizations Deliver 2.5x More Users on Nemotron 3 Ultra
From Archives Anthropic Finds AI Monitoring Can Obscure Rogue Agent Detection
From Archives OpenAI's Astra Skips Math Optimization Despite Earlier Focus
From Archives OpenAI's Math Breakthrough Relied on Our Work, Experts Say
Research & Benchmarks
Common Questions Answered
What is full-duplex audio and how does GPT-Live-1 use it differently from traditional voice bots?
Full-duplex audio allows GPT-Live-1 to listen and talk simultaneously, rather than waiting for pauses before responding like most traditional voice bots do. This capability enables more natural conversations without dead air and interruption problems that make automated calls feel unnatural.
How can developers customize GPT-Live-1 for different use cases?
Developers can mix and match backend models depending on what their task requires, allowing them to balance reasoning power, response speed, and cost. This flexibility lets teams optimize their applications for specific performance and budget needs.
What is the pricing model for GPT-Live-1 API and how does it affect development decisions?
OpenAI charges five cents per minute for GPT-Live-1 API access, which represents a significant cost consideration for developers. This pricing forces developers to make deliberate choices about whether they need full-duplex capabilities or if less expensive turn-based models would suffice for their application.
Why would a phone reservation service like Yelp benefit from using GPT-Live-1's full-duplex capabilities?
A phone reservation bot using full-duplex audio can listen while talking, which eliminates the dead air and interruption problems that make automated calls feel awkward and unnatural. This creates a more seamless user experience that better mimics human conversation patterns.
Was GPT-Live-1 already in use before the API release to developers?
Yes, the same speech model that OpenAI released as an API to developers was already running inside ChatGPT. The API release simply gave outside development teams access to this existing technology.
Further Reading
- Introducing GPT-Live - OpenAI
- How we built a realtime system for responsive voice AI in GPT-Live - OpenAI
- OpenAI launches GPT-Live voice models that listen and speak simultaneously - Reuters
- GPT-Live Pairs Full-Duplex Voice Models with a Realtime Architecture - DeepLearning.AI The Batch
- OpenAI’s GPT-Live-1 Arrives in the API at $0.05 Per Minute - Unite.AI
