GPT-Realtime-2: GPT-5-Class Reasoning for Voice
OpenAI’s new GPT-Realtime-2 model is designed to make realtime voice apps feel far closer to talking with a human. It is the company’s first voice system built with GPT-5-class reasoning, allowing it to navigate complex requests while keeping a conversation flowing. Instead of freezing when something goes wrong, the model can acknowledge errors, use short spoken preambles like “let me check that,” and call multiple tools in parallel while it works in the background. A major technical upgrade is the jump in context window from 32K to 128K tokens, enabling much longer and more coherent exchanges in voice-driven workflows. Developers can also dial reasoning up or down, from minimal for ultra-low-latency replies to xhigh when deeper analysis is required. In OpenAI’s tests, GPT-Realtime-2 significantly outperformed the previous GPT-Realtime-1.5 on audio intelligence and instruction-following benchmarks.
Realtime Voice Models and the OpenAI Audio API
The three new realtime voice models sit at the heart of an expanded OpenAI audio API aimed squarely at next-generation voice app development. GPT-Realtime-2 handles general spoken interaction and voice-to-action scenarios, where a user’s speech is converted into tasks, queries or workflows. GPT-Realtime-Translate focuses on live translation apps, while GPT-Realtime-Whisper delivers streaming, low-latency transcription. Together, these realtime voice models are built for software that responds as people speak, instead of after a long pause. OpenAI frames this as enabling three emerging patterns: voice-to-action, systems-to-voice (where software speaks back with contextual guidance), and voice-to-voice, where AI mediates conversations across languages. By integrating these capabilities through a unified Realtime API and tooling like the Codex app for Mac, OpenAI is positioning the OpenAI audio API as a foundation for richer, more dynamic voice experiences across mobile, desktop and automotive interfaces.
Live Translation Apps with GPT-Realtime-Translate
GPT-Realtime-Translate is built to keep pace with fast, multilingual conversations, opening the door to live translation apps that feel instantaneous. The model can translate speech from more than 70 input languages into 13 output languages while preserving meaning and context, even when speakers change topics mid-sentence. It can also produce realtime transcriptions alongside spoken translations, enabling use cases like bilingual customer support, cross-border sales calls, live education sessions and media localization. Early adopters are already experimenting: Deutsche Telekom is trialling the model for multilingual voice interactions, and Vimeo has demonstrated product education videos that are translated live as they play. For developers, this means one API can power both spoken interpretation and on-screen captions, reducing the need for separate pipelines. Combined with GPT-Realtime-2’s reasoning, apps can simultaneously understand intent, execute tasks and bridge language barriers in a single, fluid dialogue.
Low-Latency Transcription with GPT-Realtime-Whisper
GPT-Realtime-Whisper targets continuous speech recognition, enabling low-latency transcription that happens as people talk. Unlike batch transcription tools that process recordings after the fact, this model streams speech-to-text in realtime, making it suitable for live captions, on-the-fly meeting notes and voice-driven workflows where text needs to update as the conversation unfolds. OpenAI highlights applications in meetings, classrooms, broadcasts, customer support, healthcare, sales and recruiting—anywhere immediate processing of spoken language creates value. For developers, GPT-Realtime-Whisper can act as a front-end layer feeding text into other models or business logic, while GPT-Realtime-2 handles reasoning and response generation. This architecture lets voice apps convert audio to text, interpret it, and respond—all within a single interaction loop. Because latency is minimized, users can interrupt, correct themselves and see their words reflected quickly, making voice interfaces feel less brittle and more like natural dialogue.
Pricing, Safety and Developer Impact
OpenAI is backing its realtime voice models with clear pricing and safety frameworks aimed at production-grade voice app development. GPT-Realtime-2 is priced at USD 32 (approx. RM148) per 1 million audio input tokens and USD 64 (approx. RM296) per 1 million audio output tokens, with cached input tokens at USD 0.40 (approx. RM1.85) per 1 million. GPT-Realtime-Translate costs USD 0.034 (approx. RM0.16) per minute, while GPT-Realtime-Whisper is USD 0.017 (approx. RM0.08) per minute. On the safety side, the Realtime API includes active classifiers that monitor sessions and can halt conversations that violate harmful content policies, and developers can add custom guardrails through OpenAI’s Agents SDK. OpenAI also requires clear disclosure when users interact with AI. With companies like Zillow and Priceline already building assistants for property search and trip management, the combination of low-latency transcription, live translation and GPT-5-class reasoning is poised to make realtime voice interfaces a mainstream interaction model.