Powering Voice AI with New Voice Models in the API

OpenAI has added GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper to the Realtime API. The three models process reasoning, translation, and transcription in real time without waiting for the speaker to finish, taking voice experiences to the next level.

Voice interfaces break down into three patterns.

• voice-to-action: reasoning about a user's request to call tools and complete a task
• systems-to-voice: a service reading context and proactively guiding users by voice
• voice-to-voice: continuing a conversation across languages and contexts OpenAI explained through Zillow, Priceline, and Deutsche Telekom case studies that these patterns can work together in customer support, travel, and multilingual conversations.

GPT-Realtime-2 is the first voice model with GPT-5-level reasoning. It uses short preambles, parallel tool calling, and tool call transparency to signal that a request is being processed, and supports better error recovery, more natural interruption handling, stronger domain terminology understanding, and finer tone control. The context window has grown from 32K to 128K, and reasoning effort can be chosen among minimal, low, medium, high, xhigh, with low as the default.

Evaluation results also improved. GPT-Realtime-2 (high) scored 15.2% higher than GPT-Realtime-1.5 on Big Bench Audio, and GPT-Realtime-2 (xhigh) scored 13.8% higher on Audio MultiChallenge. Zillow reported that after prompt optimization, call success rate on a high-difficulty benchmark rose 26 points, from 69% to 95%.

GPT-Realtime-Translate translates in real time from 70+ input languages to 13 output languages, and also provides transcription while maintaining conversational pace. Deutsche Telekom uses it for multilingual customer support, Vimeo uses it to translate product training videos in real time, and BolnaAI reported that in evaluations of Hindi, Tamil, and Telugu, it achieved a 12.5% reduction in WER compared to other models, along with lower fallback rates and higher task completion rates.

GPT-Realtime-Whisper is a streaming speech-to-text model that transcribes instantly as speech is spoken. It is tailored for high-frequency voice tasks such as meeting captions, class/broadcast subtitles, notes and summaries, and customer support, healthcare, sales, and recruiting.

On the safety side, an active classifier is applied to Realtime API sessions that can halt a conversation if content is judged harmful, and developers can attach Agents SDK guardrails. The usage policy also prohibits reusing or deploying outputs for spam, deception, or harmful purposes, and requires clearly informing users that they are interacting with AI.

Pricing: GPT-Realtime-2 costs $32 per 1 million audio input tokens (cached input $0.40) and $64 per 1 million audio output tokens. GPT-Realtime-Translate costs $0.034 per minute, and GPT-Realtime-Whisper costs $0.017 per minute. All three models are available in the Realtime API.