OpenAI has taken a fresh step into the voice-tech race, unveiling three new audio models on its developer platform in a move that could change how businesses build real-time voice agents. For South African developers watching the global AI space closely, the launch signals a clear shift: ChatGPT’s maker is no longer just about text chat and transcription, but about software that can listen, respond and act while a conversation is still unfolding.
The new tools were introduced on Thursday and are now available for testing in OpenAI’s developer playground. They are called GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper, each aimed at a different part of the fast-growing voice AI market. In simple terms, OpenAI is trying to make its systems feel less like a machine and more like a capable assistant that can keep up in live conversations.
That matters because voice interfaces are moving from novelty to serious commercial product. Banks, retailers, airlines, telecoms operators and travel platforms are all under pressure to offer faster service, reduce call-centre load and improve customer experience. For companies here in South Africa, where call-centre operations remain a major part of customer support, the potential use cases are easy to see.
GPT-Realtime-2 is the headline model of the launch. OpenAI says it is built to handle more complex requests, work with tools, manage interruptions and keep track of context over longer voice sessions. That is a crucial feature for any agent designed to assist a customer from start to finish, rather than simply answering one question at a time.
In practical terms, that could mean a voice system capable of handling a booking, checking availability, confirming details and resolving a follow-up query without losing the thread. For businesses, that kind of continuity is what separates a gimmick from a genuinely useful AI agent.
The second model, GPT-Realtime-Translate, is aimed at translation. OpenAI says it can translate speech from more than 70 languages into 13 output languages, opening the door to customer service, training and education tools that can work across borders and language groups. That will be watched closely in markets like South Africa, where multilingual communication is a daily reality and where companies often struggle to serve customers in their preferred language at scale.
The third model, GPT-Realtime-Whisper, is designed for live speech-to-text. It can generate captions, meeting notes and workflow updates as a person speaks, making it useful in situations where speed and accuracy matter. From boardrooms to classrooms, and from remote work to compliance-heavy industries, real-time transcription is becoming a valuable layer in AI-powered productivity tools.
Customers already testing the new models include Zillow, Priceline and Deutsche Telekom. That is telling. These are not experimental hobby projects; they are large, established businesses that depend on scale, reliability and customer trust. Their involvement suggests OpenAI is positioning the models for commercial deployment rather than simple demonstration.
For South African readers, the timing is significant. Local companies have been increasingly open to AI tools that can cut costs and improve service delivery, but many remain cautious about accuracy, bias, data privacy and integration challenges. New audio APIs like these are likely to attract interest from fintechs, insurers, property platforms and telecoms groups looking to automate more of the front line without sounding robotic.
OpenAI’s new audio models could reshape real-time voice agents
The arrival of OpenAI’s new audio models also highlights how quickly the market for voice AI is maturing. Until recently, most companies treated speech technology as two separate tasks: speech-to-text on one side, and text-to-speech or chat on the other. OpenAI is now pushing for something more seamless, where the AI can handle the whole interaction in real time.
That shift could be especially important in customer support environments. A good voice agent needs to understand interruptions, maintain context, and continue helping even when a caller changes direction mid-sentence. According to OpenAI, GPT-Realtime-2 is designed with exactly that in mind, which may make it more suitable for call-centre style workflows than older, narrower models.
There is also a broader strategic angle here. By releasing audio models through its developer platform, OpenAI is giving businesses tools they can build into their own applications, rather than forcing them to use a generic chatbot interface. That creates opportunities for product teams to design very specific voice experiences around bookings, support, education, logistics and internal admin.
Pricing will be another point of interest for developers and procurement teams. OpenAI says GPT-Realtime-2 starts at US$32 per million audio input tokens. GPT-Realtime-Translate is priced at $0.034 per minute, while GPT-Realtime-Whisper comes in at $0.017 per minute. Those numbers will be scrutinised closely by businesses trying to balance innovation with cost control.
For smaller firms, the pricing may still be a hurdle, especially if voice AI is used at scale. But for larger enterprises, the real question is whether the models can reduce labour costs, improve response times and create a better customer experience than current systems. If they can, the economics may quickly justify the spend.
As we’ve reported before on the global AI race, OpenAI is under increasing pressure to keep expanding beyond text. Voice is one of the most obvious battlegrounds, because it offers a more natural way for people to interact with software. It is also one of the hardest areas to get right, since real conversations are messy, fast and often unpredictable.
For that reason, the launch of these models is about more than a product update. It is a sign that OpenAI wants to be at the centre of how the next generation of voice-based AI agents is built and deployed. And for South African businesses keeping an eye on what’s coming next, these tools are a reminder that the future of customer service may soon sound a lot more human.