OpenAI launches GPT-Realtime models with 70-language voice input support

This release shifts voice AI from simple call-and-response to continuous streaming agents capable of real-time reasoning and translation. For the streaming industry, these models provide a standardized framework for automated live captioning and multilingual accessibility without the latency of traditional transcription-then-synthesis pipelines. By embedding 70-language support directly into the API, OpenAI is targeting global enterprise services like travel and technical support that require high-concurrency, low-lag verbal interaction. Watch for developer adoption rates in live-streaming apps where instantaneous, accurate translation is the primary barrier to international content distribution.

OpenAI has moved its Realtime API from beta to general availability with significant architectural changes. The GA release introduced the gpt-realtime model alongside new features including image input, async function calling, MCP support, and EU data residency, replacing the earlier beta interface that required a special header. Developers building voice agents now connect via WebRTC or WebSocket using ephemeral client secrets, with sessions capped at 60 minutes. The API supports both conversational speech-to-speech interactions and a dedicated transcription mode for workflows where no model response is expected, making it directly applicable to live captioning pipelines.

For streaming and broadcast use cases, OpenAI has separated translation into its own dedicated endpoint. The gpt-realtime-translate model connects to a distinct /v1/realtime/translations path and streams translated audio plus transcript deltas while the speaker is still talking, targeting live interpretation, multilingual calls, broadcasts, meetings, and video rooms. Unlike the standard voice-agent session, the translation session does not use a conversation or response lifecycle and does not call tools. It produces continuous translated audio output, which positions it as a drop-in replacement for traditional simultaneous interpretation in live events and multilingual streaming scenarios.

The broader Realtime API architecture gives developers granular control over session behavior. OpenAI's documentation specifies that the Realtime API supports barge-in, low first-audio latency, natural turn taking, and realtime tool use within a single session, with the Agents SDK providing a higher-level abstraction for browser applications. For teams building live captioning or translation features into streaming platforms, the key decision point is whether to use gpt-realtime-2.1 in transcription mode for caption-only workflows or gpt-realtime-translate when the output must be spoken audio in another language. The 60-minute session limit and WebRTC transport layer suggest the API is optimized for live, synchronous interactions rather than long-form batch processing.