OpenAI Releases GPT-Realtime-2.1 Voice Models With Lower Latency

OpenAI released GPT-Realtime-2.1 and a distilled mini variant on July 6, 2026, cutting p95 latency by at least 25% through improved caching and enhancing alphanumeric recognition, noise handling, and interruption behavior. The update targets voice-agent builders with speech-to-speech support, reasoning control, and tool use, priced at $4/$24 per million text tokens and $32/$64 per million audio tokens, signaling rapid iteration on production voice infrastructure.

OpenAI's point release for its Realtime voice family is a reminder that the API-level battle for production voice agents is now being fought on latency and reliability metrics, not just raw model capability. OpenAI shipped gpt-realtime-2.1 and a distilled gpt-realtime-2.1-mini variant, according to the official model documentation and an OpenAI Developer Community announcement posted July 6, 2026. The update targets voice-agent builders directly: OpenAI says p95 latency across Realtime voice models dropped by at least 25%, driven largely by improved caching, alongside better alphanumeric recognition, noise and silence handling, and more natural interruption behavior. Both models keep speech-to-speech support with configurable reasoning effort, instruction following, and tool use, with the full model priced at $4/$24 per million text tokens and $32/$64 per million audio tokens. For teams building call-center bots, IVR replacements, or voice copilots, the release signals that OpenAI is iterating on production voice infrastructure roughly every two months rather than treating Realtime as a settled product line.