OpenAI realtime-2.1 cuts voice latency 25%
Two new Realtime API models cut p95 latency by at least 25% and add configurable reasoning effort and tool use, available in the OpenAI API now.
OpenAI released gpt-realtime-2.1 and gpt-realtime-2.1-mini on July 6. Both replace the previous realtime family in the API and are available now for voice and multimodal agent applications.
The headline improvement is a 25% reduction in p95 latency across Realtime voice models, achieved through improved caching. The models also handle background noise, silence detection, and call interruptions more reliably, and support configurable reasoning effort and tool use.
Pricing
| Model | Text input | Audio input | Audio output |
|---|---|---|---|
| gpt-realtime-2.1 | $4.00/M | $32.00/M | $64.00/M |
| gpt-realtime-2.1-mini | $0.60/M | $10.00/M | $20.00/M |
The mini variant costs roughly one-seventh the full model on audio. Cached input pricing is lower for both.
What this means for builders
The separation into two tiers creates a routing decision. gpt-realtime-2.1 is the right choice for agents that use tools or need to follow complex instructions mid-call. gpt-realtime-2.1-mini handles simpler flows (answering questions, routing calls, collecting structured data) at substantially lower cost.
The 25% latency cut matters most for conversational turn-taking. Human phone calls tolerate roughly 200ms of delay before responses feel unnatural. Shaving latency at the 95th percentile means fewer calls hit that threshold during connection hiccups or complex tool lookups.
The guide on putting an AI agent on your phone lines covers the architecture decisions that precede model selection. If you have existing voice agents on the prior Realtime API, the migration path is a model name change with no other API changes required.