OpenAI gpt-realtime-2.1

OpenAI gpt-realtime-2.1 is a native speech-to-speech model released July 6, 2026, available via WebRTC, WebSocket, and direct SIP trunking. It processes bidirectional audio streams within a stateful session carrying up to 128K tokens of dialogue history, enabling long-form conversations without context truncation. The July 2026 release reduced p95 voice latency by at least 25 percent compared to gpt-4o-realtime-preview, bringing typical round-trip latency to approximately 320ms. Audio tokens are priced at $32 per million for input and $64 per million for output, equating to roughly $0.10 per minute of active conversation at standard rates. Prompt caching reduces input costs by 80x, lowering sustained-call pricing significantly for agents with stable system prompts. The API provides eleven preset voices, native function calling mid-conversation, and barge-in interruption handling. No separate STT or TTS swap is required or supported: the model is an end-to-end architecture. Enterprise contact center teams benefit from native SIP trunking, a 99.9 percent contractual uptime SLA on enterprise plans, and scalable concurrency negotiated beyond the standard 100-session limit. The model's accuracy on streaming audio and voice naturalness position it as the premium choice for production voice agents requiring deep OpenAI ecosystem integration.