A voice AI that can actually reason is a different product category from a voice AI that simply responds. That distinction matters because most businesses that tried voice automation in the past few years ran into the same wall: the system handled simple, scripted scenarios well, and fell apart the moment a customer asked something slightly unusual.
OpenAI launched GPT-Realtime-2 on May 7, 2026, alongside two companion models: GPT-Realtime-Translate, which handles live voice translation across more than 70 languages, and GPT-Realtime-Whisper, a streaming transcription model. The three together represent OpenAI taking the Realtime API out of beta and positioning it as production infrastructure.
GPT-Realtime-2 is the significant one. Previous voice models processed speech in a simple loop: you speak, the model responds, reset. The new model can reason while the conversation continues, call multiple external systems at the same time during a call, and maintain context across a much longer exchange. The context window is now 128,000 tokens, four times larger than before. In practical terms: it will not lose track of something said at the start of a 30-minute call.
Developers can also set how hard the model thinks. A simple appointment booking call does not need deep reasoning, and you can set it accordingly to save cost and reduce response time. A complex insurance claim call can be set to higher reasoning effort. That dial is new, and it matters for cost control.
Pricing sits at $32 per million audio input tokens and $64 per million audio output tokens. The translation model runs at $0.034 per minute, and the transcription model at $0.017 per minute. These are developer API prices, not consumer subscription fees. Any company building a voice product pays per use.
The broader context is worth understanding. Gartner forecasts that conversational AI will handle roughly 70% of enterprise customer support interactions by 2027. That projection was made before this release. The gap between what voice AI could do and what a human support agent could do was the main reason adoption was slow. GPT-Realtime-2 narrows that gap substantially, particularly for calls that require judgment rather than just retrieval.
Early deployments are already visible. Zillow is using the model for client calls covering home valuations and financing scenarios, the kind of conversation that previously required a knowledgeable human because the questions are situational and unpredictable.
For business operators, the practical question is: do you run voice-facing processes? Customer service lines, intake calls, appointment booking, order updates, insurance queries, field service dispatch. If yes, the economics of keeping all of those calls human-staffed are becoming harder to defend, not because the technology was previously good enough, but because it now is.
This release also includes live translation across more than 70 input languages into 13 output languages. For businesses that operate across multiple countries with phone-based customer interaction, that is a separate and immediate opportunity.
The consumer-facing ChatGPT app does not yet reflect these improvements. OpenAI has indicated updates are coming, but what launched in May is a developer tool. The products your customers interact with will be built on top of it over the next 12 to 18 months.