OpenAI releases GPT-Realtime-2.1 and a mini version for voice agents in its API

OpenAI has added two new models from its Realtime line to its API: GPT-Realtime-2.1 and the lighter GPT-Realtime-2.1-mini. Both are designed for voice agents that need low response latency.

What's new in the Realtime line

GPT-Realtime-2.1-mini is a compact reasoning model built for voice scenarios. Its price matches that of the earlier gpt-realtime-mini, so moving to the new version does not raise costs for developers who already used the previous mini model. OpenAI's Realtime line was built from the start for tasks where text models that generate an intermediate response are a poor fit: voice assistants, conversational support interfaces, and agents that must react to a user's remarks almost without a pause.

Releasing two models at once, the flagship GPT-Realtime-2.1 and the lighter mini version, gives developers a choice between a more powerful reasoning model and a cheaper option for cases where speed and cost matter more than depth of reasoning in each reply.

• Two new models: GPT-Realtime-2.1 and GPT-Realtime-2.1-mini
• GPT-Realtime-2.1-mini is a compact reasoning model for voice
• The mini version's price is comparable to the previous gpt-realtime-mini
• p95 latency is down by at least 25% thanks to improved caching
• The models are accessed over the WebRTC protocol

Why p95 latency matters

For voice agents that must answer users in real time, latency is one of the key quality metrics: the higher it is, the more mechanical and unnatural a conversation feels. The p95 metric shows the latency that 95% of requests do not exceed, so it reflects how the system behaves not in the best case but in a typical, near-worst case. A reduction of at least 25% through improved caching means the system now delivers the vast majority of voice responses noticeably faster than before, without any change to the reasoning model itself.

How to connect to the new models

OpenAI has kept WebRTC as the connection method for the Realtime line, the same one used for previous versions. This lets developers who have already built voice agents on the Realtime API switch to the new models without changing the transport protocol.

What this means

The Realtime update shows that OpenAI keeps making targeted improvements to its voice-agent infrastructure, cutting latency and lowering the cost of access to reasoning models for voice, rather than only shipping more powerful but more expensive versions.