- Published on
OpenAI Releases GPT-Realtime-2.1 and GPT-Realtime-2.1-mini for Low-Latency Voice Agents in the API
- Authors
- Name
- DeepHorus
OpenAI has released two new realtime voice models in its API: gpt-realtime-2.1 and gpt-realtime-2.1-mini. Both target low-latency voice and multimodal experiences, with the mini model being a notable addition for its reasoning capabilities. The release aims to enhance voice agents by reducing latency and improving interactive quality.
The Realtime API processes and generates audio through a single model, avoiding the need to chain separate speech-to-text and text-to-speech systems. This design reduces latency and preserves nuance in speech. The mini model, gpt-realtime-2.1-mini, is a reasoning model that can think internally before speaking, supporting tool use and function calling within live voice interactions.
Key improvements include a reduction in p95 latency by at least 25% across realtime voice models, driven by improved caching. Pricing remains competitive: gpt-realtime-2.1-mini audio output costs $20 per 1M tokens, while the full gpt-realtime-2.1 charges $64. Cached audio input for the mini drops to $0.30 per 1M tokens, compared to $10 for fresh input, benefiting long sessions.
Reasoning and tool use address a common issue in voice agents where stalls during tool calls lead to user interruptions. The model can now say 'I'll check that order now' before acting, keeping multi-step tasks coherent. Reasoning effort is configurable across levels from minimal to xhigh, with low being the default for simple turns.
OpenAI positions gpt-realtime-2.1-mini as the faster, more cost-efficient option for realtime voice. Use cases include customer support triage, where the mini model reasons about billing issues and narrates steps, or appointment scheduling with precise date recognition. In-app voice assistants can stream audio over WebRTC with low latency.
The release is expected to drive wider adoption of voice agents by balancing cost and capability. Developers can trade some reasoning depth for lower latency and cost, while maintaining interactive quality. As caching and latency improvements mature, realtime voice applications may become more prevalent in customer service and mobile interfaces.
