OpenAI has launched three new audio models in the API , GPT-Realtime-2 for voice reasoning, GPT-Realtime-Translate for live translation across 70 input and 13 output languages, and GPT-Realtime-Whisper for streaming speech-to-text , positioning voice as a developer platform for natural, interruptible agents that handle complex requests, tools, and multilingual workflows without traditional STT-LLM-TTS pipelines.
The models represent a fundamental shift from chained processing to native audio handling. GPT-Realtime-2 processes speech end-to-end, preserving tone, emotion, and rhythm that text intermediaries flatten. It supports GPT-5-class reasoning, asynchronous function calling, and natural interruptions. GPT-Realtime-Translate keeps pace with speakers, handling 70+ input languages to 13 output languages. GPT-Realtime-Whisper transcribes live as people talk. New voices Marin and Cedar add naturalness. All models are available through the existing API with pricing at $32 per million input audio minutes and $64 per million output.
The technical architecture eliminates pipeline latency. Traditional voice apps chain Whisper for STT, GPT for reasoning, and TTS for output, adding 500 to 1,000 milliseconds per turn. The new models process audio directly, cutting latency while maintaining conversational flow. Developers can instruct TTS to adopt specific styles, like "sympathetic customer service agent," for customised experiences. Asynchronous tools allow the model to speak while long-running functions resolve, making voice agents feel responsive even for complex tasks. The SDK supports building full voice pipelines with modality control and glitch-free output.
The voice API economics favour high-volume applications. Pricing remains premium compared to text, but latency reductions and native capabilities lower total system cost for realtime use cases. Call centers can replace human agents with multilingual voice bots that handle interruptions and tool calls. Tutoring apps can provide instant feedback in any language. Healthcare intake can transcribe and reason over patient speech. Personal assistants can run hands-free workflows across devices. The realtime nature makes voice viable for safety-critical applications like emergency response or in-car navigation where delays matter.
For SF readers, the launch opens voice as a new startup surface area. Voice-native apps have lagged text due to pipeline complexity and poor economics. These models make building interruptible, multilingual agents straightforward. Startups can embed voice into consumer products like fitness coaching, language learning, or companion apps. Enterprise use cases include automated customer support, virtual receptionists, and compliance call monitoring. The API-first approach lowers barriers compared to proprietary voice hardware or SDKs.
OpenAI builds a moat through multimodal integration. GPT-Realtime-2 supports text and image inputs alongside audio, enabling vision-enabled voice agents. The ecosystem of OpenAI libraries, tracing, and deployment tools gives developers a complete stack. Independent voice AI startups face platform risk: superior models from OpenAI make differentiation hard. The winners will build applications that leverage realtime voice in verticals where domain data and workflows create defensibility, not standalone voice models.
The call center angle shows the economics most clearly. Traditional centres cost $35,000 to $60,000 per agent annually. A voice agent handling 80 percent of calls at $0.10 per minute saves $500,000 per agent per year at scale. Multilingual capabilities eliminate translation overhead. Tutoring scales to millions with personalised feedback. Healthcare intake reduces clinician time on history-taking. The savings compound as models improve.
Also read: ChatGPT's Trusted Contact feature turns AI assistants into safety infrastructure for the first time • DeepMind's AlphaEvolve uses Gemini to optimise its own infrastructure and rediscover mathematics • Microsoft's VS Code Copilot co-author blunder reveals the governance tension in AI developer tools
