OpenAI Releases gpt-realtime-2.1 With 25% Lower Latency for API Voice Agents

My take: OpenAI today released gpt-realtime-2.1 and gpt-realtime-2.1-mini, updates to its voice API models with at least 25% lower p95 latency. The improvement comes from inference-layer caching optimizations, paired with better alphanumeric recognition, improved silence and noise handling, and support for configurable reasoning effort and tool use directly within a voice session.

For anyone building voice AI applications, latency is the metric that most affects user perception. A response that takes more than 800 milliseconds feels mechanical; one that arrives before 400 starts to feel like a real conversation. A 25% reduction in p95 latency, the percentile that tracks the slowest cases rather than the average, means the spikes that ruin user experience become less frequent.

I use voice agents in my own projects, and this kind of update, though quiet, has real impact on the end user experience. Are your current voice workflows running on the latest API models, or have they been on an older version for months while newer options went live?