OpenAI releases GPT-Realtime-2.1 and GPT-Realtime-2.1-mini for low-latency voice agents

OpenAI has released two new Realtime models in its API, named gpt-realtime-2.1 and gpt-realtime-2.1-mini. Both models target low-latency voice and multimodal experiences. The mini model is a notable part of this release, as it is a mini reasoning model for realtime voice that ships at the same cost as the earlier gpt-realtime-mini.

gpt-realtime-2.1-mini is a mini reasoning model for realtime voice interactions that responds to audio and text inputs over a live connection. OpenAI positions it as the faster, more cost-efficient option in the lineup. The Realtime API processes and generates audio through a single model, avoiding chaining separate speech-to-text and text-to-speech systems.

This single-model design reduces latency and preserves nuance in speech. The main capability of these models is reasoning, which means the model can think internally before it speaks. The mini tier also supports tool use, or function calling, through the Realtime API.

Together, these let the mini model plan a step, call a function, and then answer. The larger sibling is gpt-realtime-2.1, which updates GPT-Realtime-2 with improved alphanumeric recognition, silence and noise handling, and interruption behavior. It supports speech-to-speech with configurable reasoning effort, instruction following, and tool use.

When choosing between the two models, use gpt-realtime-2.1 when you want the strongest realtime reasoning, tool use, instruction following, and voice-agent behavior. Use gpt-realtime-2.1-mini when you want a faster, more cost-efficient option. Voice agents often stall during tool calls, but reasoning and a spoken preamble fix this pattern.

The model can say 'I'll check that order now' before acting, keeping talking while it works through a request. This behavior keeps multi-step voice tasks coherent. Reasoning effort is configurable across levels, from minimal to xhigh.

Low is the default and keeps latency down for simple turns. Higher effort increases latency and output token usage. OpenAI advises starting low for most production voice agents.

The 95th-percentile response time, or p95 latency, has been reduced by at least 25% across Realtime voice models, driven by improved caching. Caching also lowers cost, not just latency, with cached input tokens billed at a steep discount. Pricing is per 1M tokens, split by text, audio, and image.

The mini model keeps the previous mini rate while adding reasoning. The mini audio output rate is $20.00 per 1M, while the full gpt-realtime-2.1 charges $64.00 for the same. To use the Realtime API, browser clients connect over WebRTC, while server media pipelines use WebSockets, and telephony uses SIP.