OpenAI introduced three realtime voice models designed to enable developers to build voice applications with reasoning, translation, and transcription capabilities. The new models, released May 7, 2026, expand the API offerings for creating voice experiences that respond intelligently and operate in real time.
The three realtime voice models include GPT-Realtime-2, which combines voice interaction with reasoning capabilities similar to GPT-5-class models. GPT-Realtime-Translate provides live translation across 70 input languages and 13 output languages. GPT-Realtime-Whisper delivers streaming speech-to-text transcription with low latency.
GPT-Realtime-2: Voice with Reasoning
GPT-Realtime-2 processes voice conversations while reasoning through requests, calling tools, and handling interruptions. The model supports context windows of 128K tokens, up from 32K in earlier versions. Developers can adjust reasoning effort across five levels: minimal, low, medium, high, and xhigh. Performance evaluations showed GPT-Realtime-2 scoring 15.2% higher on Big Bench Audio for audio intelligence compared to GPT-Realtime-1.5, and 13.8% higher on Audio MultiChallenge for instruction following.
The model includes features for more natural interactions. Preambles let developers add short phrases like “let me check that” before main responses. Parallel tool calls enable the model to execute multiple actions simultaneously while announcing them to users. Stronger recovery behavior allows graceful handling of errors and interruptions.
Live Translation Across Languages
GPT-Realtime-Translate enables developers to build applications where users speak in their preferred language while receiving real-time translations. The model handles 70 input languages and translates into 13 output languages. Testing by BolnaAI across Hindi, Tamil, and Telugu showed 12.5% lower Word Error Rates compared to other models tested, along with lower fallback rates and natural conversation latency.
Streaming Transcription with Low Latency
GPT-Realtime-Whisper provides streaming speech-to-text transcription designed for low-latency speech recognition. The model transcribes audio as people speak, enabling live captions, meeting notes, and voice agents that understand users continuously. The transcription capability supports business workflows in customer support, healthcare, sales, recruiting, and other high-volume spoken interactions.
Developer Use Cases and Partners
Zillow is building a voice assistant using GPT-Realtime-2 to handle real estate requests, listening and reasoning through complex queries like finding homes within budget constraints while avoiding specific locations and scheduling tours. The company reported a 26-point lift in call success rate on adversarial benchmarks after prompt optimization, reaching 95% success versus 69% previously.
Deutsche Telekom is testing GPT-Realtime-Translate for multilingual customer support where agents and customers can converse in their preferred languages. Priceline is developing voice interfaces for travel management, enabling users to search flights and hotels conversationally, handle reservation changes, and receive real-time updates on airport conditions.
Pricing and Safety
GPT-Realtime-2 is priced at $32 per million audio input tokens with $0.40 per million for cached input tokens, and $64 per million audio output tokens. GPT-Realtime-Translate costs $0.034 per minute. GPT-Realtime-Whisper costs $0.017 per minute. The Realtime API incorporates safety classifiers that can halt conversations violating content guidelines. Developers can add custom safety guardrails using the Agents SDK. The API supports cybersecurity requirements including EU Data Residency for European applications.
Developers can test the models in the Playground or use Codex to integrate GPT-Realtime-2 into existing applications. The new realtime voice models represent a shift from simple voice call-and-response toward voice interfaces capable of listening, reasoning, translating, transcribing, and taking action throughout a conversation.
Source: OpenAI
