OpenAI adds GPT-Realtime-2, Translate and Whisper to its API for voice apps

On May 7, 2026, OpenAI introduced three realtime voice models in its API: GPT-Realtime-2 for conversation and actions, GPT-Realtime-Translate for live translation, and GPT-Realtime-Whisper for streaming transcription. The company is clearly pushing voice interfaces away from a "reply to one utterance" mode toward an assistant that can listen, reason, use tools and stay in the conversation.

Three models at once

The core idea of the release is simple: voice in apps should work not as a nice add-on but as a full interface. OpenAI says developers increasingly build three types of scenarios: voice-to-action, where the user states a task by voice and the system carries out the actions itself; systems-to-voice, where software proactively tells the person what is happening; and voice-to-voice, where AI helps people hold a conversation across different languages. The new lineup was built for this set of scenarios.

• GPT-Realtime-2 is a voice model with GPT-5-class reasoning, tool-calling support and a longer context.
• GPT-Realtime-Translate translates speech in real time from more than 70 input languages into 13 output languages with almost no pauses.
• GPT-Realtime-Whisper is streaming transcription that writes text as the person speaks, not after the phrase ends.
• Pricing was announced at the same time: GPT-Realtime-2 costs $32 per 1M audio input tokens and $64 per 1M output tokens, Translate costs $0.034 per minute, and Whisper costs $0.017 per minute.

All three models are already available through the Realtime API, and you can test them in the Playground. This matters: OpenAI is not showing a distant concept but shipping a ready set of tools for teams building support, voice agents, real-time translation, meeting notes and other products involving live speech. For the market, it is a signal that the tools are ready not only for demos but also for pilots.

What improved in dialogue

The most noticeable update is GPT-Realtime-2. The model can insert short status phrases such as "let me check" so the user understands the system is busy with a task. It can call several tools in parallel, narrate its actions aloud, recover better from errors and interruptions, and sustain noticeably longer scenarios: the context window grew from 32K to 128K. For production this matters far more than a "pleasant voice", because real assistants usually break on long chains of steps.

OpenAI also stresses the model's controllability. Developers can choose a reasoning level from minimal to xhigh, balancing latency against answer quality. Understanding of terms, proper names and specialized vocabulary, such as medical terms, has also improved. In internal evaluations, GPT-Realtime-2 in high mode scored 15.2% higher than GPT-Realtime-1.5 on Big Bench Audio, and in xhigh mode it was 13.8% better on Audio MultiChallenge, which measures instruction following in conversation.

"After tuning the prompts, we saw call success rise from 69% to 95%," is how

Zillow describes its early tests of GPT-Realtime-2.

Translation and transcription

The second model, GPT-Realtime-Translate, targets live multilingual conversation. It translates speech as the conversation unfolds, keeping the speaker's pace and meaning even when people speak with accents, jump between topics or use industry jargon. OpenAI separately cites use cases in support, cross-border sales, education, events, media and creator platforms. Deutsche Telekom is testing the model in multilingual customer support, and Vimeo demonstrates a scenario where a training video is translated as it plays.

The third model, GPT-Realtime-Whisper, handles a more down-to-earth but very much in-demand task: turning speech into text quickly. OpenAI positions it as a foundation for subtitles, meeting notes, transcription of classes and broadcasts, and for voice agents that need to continuously understand what a person is saying.

The company also points to safeguards: active classifiers run in the Realtime API, some sessions may be stopped if rules are violated, and developers must clearly tell users when they are talking to an AI.

What this means

OpenAI is trying to claim not only the chat-model market but also the base layer for voice products. If quality and latency really match the stated metrics, the company gets a strong position in call centers, travel services, educational platforms and enterprise assistants, where what matters is not flashy demos but stable conversation, translation without pauses, and text that appears at the very moment the person speaks.