Reasoning and tool use are coming to OpenAI's Realtime mini range.

OpenAI's GPT-Realtime-2.1-mini API brings reasoning and tool use to its lightweight voice model at no extra cost, while reducing p95 latency by 25%.

OpenAI's real-time voice model family expands with GPT-Realtime-2.1-mini, now available via API. Its unique feature is bringing capabilities previously limited to heavier versions into a lightweight format: reasoning, which allows the model to chain multiple steps of thought before responding, and tool use, which enables it to trigger external functions during a conversation, such as querying a database or initiating an action. All of this is billed at the same rate as GPT-Realtime-mini, with no additional cost for these enhancements.

In parallel, the publisher reports a gain affecting all of its real-time voice models: p95 latency, the delay experienced by the slowest requests, has decreased by at least 25% thanks to a revamped caching system. Both developments converge on the same sensitive point for voice interfaces, reaction speed, which determines the naturalness of a spoken exchange with a machine and separates a fluid conversation from a dialogue chopped up by waiting times.