OpenAI launches GPT-Realtime-2, Translate, and Whisper APIs
OpenAI released three new streaming audio models in its Realtime API: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. The company says GPT-Realtime-2 adds better reasoning, longer context, tool use, and interruption handling for voice agents, while the other two models handle live translation and transcription.
OpenAI has launched three new streaming audio models in the Realtime API: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. The company says the release is aimed at developers building voice agents that can listen, respond, use tools, and keep up with live conversation.
According to OpenAI, GPT-Realtime-2 is its most capable voice model so far. The company described it as bringing “GPT-5-class reasoning” to real-time speech systems, with better handling of interruptions, more natural turn-taking, and stronger recovery when something goes wrong.
⚡ New to this?
This matters because voice AI is moving beyond simple speech-to-text. A voice model is the part that listens and speaks back; transcription turns speech into text, and translation converts speech from one language to another.
For non-experts, the big change is that these systems are getting better at holding a live conversation while also using tools like calendars or databases. That can affect customer support, meetings, accessibility, and live translation tools.
🦞 OpenClaw angle
If you build voice agents, treat the model as part of a real-time system, not just a chatbot. Add explicit states for interruption handling, tool-call announcements, and recovery messages so the agent does not sound broken when it pauses or checks something.
Use the new reasoning controls to separate fast replies from slower, more careful ones. If your workflow depends on long context, plan for session memory and prompt hygiene so the agent can keep names, entities, and task state straight over time.
For multilingual systems, test GPT-Realtime-Translate against your actual language mix instead of assuming general translation quality is enough. For internal note-taking or captions, wire GPT-Realtime-Whisper into a streaming pipeline so you can consume text as it arrives, not after the call ends.
The model is also designed to make voice agents easier to follow. OpenAI said developers can use short preambles before a main answer, such as “let me check that,” and can make tool use audible, for example by saying “checking your calendar” while the agent is working.
OpenAI said GPT-Realtime-2 can call multiple tools at once and is better at maintaining state across longer conversations. The company also said it has improved support for specialized vocabulary, proper nouns, and healthcare terms, along with more control over tone and delivery.
Another change is the larger context window. According to reporting cited in the source, GPT-Realtime-2 increases context from 32K to 128K tokens, which gives voice agents more room to carry conversation history and task details through a session.
OpenAI also added adjustable reasoning levels for the model. The available settings are minimal, low, medium, high, and xhigh, with low as the default, according to the source article’s summary of the release.
The two companion models handle translation and transcription. GPT-Realtime-Translate supports live speech translation from more than 70 input languages into 13 output languages, according to OpenAI. GPT-Realtime-Whisper provides streaming transcription as speech is produced, which OpenAI said can be used for captions, notes, and continuous speech understanding.
The company said the models are available now in the Realtime API. ChatGPT’s own voice mode has not been upgraded yet, and OpenAI has only said that those changes are still coming.
The launch was accompanied by several product demos and early integrations. OpenAI said companies including Glean, Vimeo, and Genspark have already used the new models in voice products or internal evaluations. Glean reported a 42.9% relative increase in helpfulness in its internal tests, while Genspark said its Call for Me Agent saw a 26% higher effective conversation rate after moving to GPT-Realtime-2.
Independent benchmark results cited in the source also pointed to stronger performance. Artificial Analysis reported a 96.6% score on Big Bench Audio for the high-reasoning variant, while Scale AI said GPT-Realtime-2 reached the top spot on its Audio MultiChallenge speech-to-speech leaderboard.
OpenAI is positioning the new models as building blocks for real-time applications rather than as a consumer voice assistant update. The company’s release centers on live voice interaction, translation, and transcription for developers using the Realtime API.
