From Demo Bots to Real-Time Voice AI Infrastructure
OpenAI’s new GPT-Realtime models mark a shift from showcase voice assistants to infrastructure for production-grade real-time voice AI. The company has introduced three models through its Realtime API: GPT-Realtime-2 for advanced conversational AI reasoning, GPT-Realtime-Translate for multilingual speech, and GPT-Realtime-Whisper for live voice transcription. Rather than relying on one monolithic model to handle every task, the stack separates capture, translation, and reasoning into distinct lanes. This design targets live systems that must keep speaking, survive interruptions, and call external tools without losing context. OpenAI frames the lineup as the backbone for call flows, support desks, travel agents, and media workflows that need continuous dialogue, not one-off answers. By decoupling capabilities, teams can mix and match depth, latency, and cost according to each turn in a conversation, instead of over-provisioning a single heavy model for every utterance.
GPT-Realtime-2: GPT-5-Class Reasoning in Live Conversations
GPT-Realtime-2 sits at the reasoning tier of the stack, bringing what OpenAI describes as “GPT-5 class reasoning” into live conversations. It is designed to manage tools, handle multi-turn logic, and maintain context when users interrupt, change topics, or trigger external workflows. Traditionally, developers have compensated for model limits with session resets and state compression, constantly reconstructing context when calls get long or complex. With GPT-Realtime-2, more of that orchestration is pushed into the model layer itself. The goal is a voice agent that can fetch account data, answer follow-up questions, switch tools, and then return to the original thread without losing state. For developers, this reduces the glue code needed to keep conversations coherent. It also supports richer conversational AI reasoning in domains like customer support, travel bookings, and complex troubleshooting, where each turn builds on a long chain of prior interactions.
Dedicated Models for Translation and Live Voice Transcription
Alongside the reasoning tier, OpenAI has separated translation and transcription into their own GPT-Realtime models. GPT-Realtime-Translate focuses on speech translation, handling 70 input languages and producing real-time output in 13 languages. This lets developers plug a specialized voice translation API into workflows for customer support, travel assistance, and media localization without overloading the main assistant. GPT-Realtime-Whisper, meanwhile, is a low-latency streaming speech-to-text system tuned for live voice transcription. By isolating transcription from reasoning, teams no longer need to route every spoken word through the most expensive model just to capture text reliably. This structure supports multilingual call centers, interactive kiosks, and media tools where accurate, continuous transcription is critical, but deep reasoning is only needed at specific decision points. The result is a more flexible pipeline that treats capture, translation, and understanding as modular services rather than a single black box.
Why a Split Architecture Matters for Developers and Enterprises
The split architecture behind the GPT-Realtime models gives developers more control over both performance and cost. In many enterprise voice systems, a single bundled model becomes a bottleneck, handling everything from raw speech capture to complex reasoning. That design can create latency spikes and opaque failure modes, especially during long, multilingual, or tool-heavy calls. OpenAI’s approach separates these concerns so teams can tune each lane independently. For example, a support flow may rely on fast, accurate transcription and lightweight routing for most of a call, invoking GPT-Realtime-2 only when a decision-heavy step truly benefits from deeper reasoning. The separation also simplifies debugging: if transcription remains accurate while reasoning slows, teams can narrow issues to the reasoning tier instead of replacing the entire stack. Procurement and engineering teams gain clearer benchmarks, testing where transcription, translation, or reasoning breaks first under real-world workloads rather than treating the voice system as an indivisible unit.
Cost, Competition, and Early Real-World Workloads
Pricing signals that OpenAI is targeting production deployments rather than experimental demos. GPT-Realtime-2 is listed at USD 32 (approx. RM150) per 1 million audio input tokens and USD 64 (approx. RM300) per 1 million audio output tokens. GPT-Realtime-Translate is priced at USD 0.034 (approx. RM0.16) per minute, while GPT-Realtime-Whisper costs USD 0.017 (approx. RM0.08) per minute. This granular pricing lets teams model costs separately for speech-heavy versus reasoning-heavy workflows. OpenAI is positioning the stack against rival real-time voice AI offerings from players such as Microsoft and xAI, but is tying its pitch to specific deployment categories. Early use cases span Zillow’s property interactions, Deutsche Telekom’s multilingual support, Priceline’s travel assistance, and Vimeo’s media translation. These workloads stress-test conversational AI reasoning, live voice transcription, and translation together. Buyers can now evaluate whether every turn requires the full reasoning tier—or whether a modular, lane-based approach delivers the right blend of responsiveness, reliability, and economics.