OpenAI has released two updated additions to its Realtime API lineup, aimed at developers building voice agents and other speech-to-speech applications, according to the company’s July 6 announcement.
Faster responses
The headline change is speed: OpenAI says the update cuts p95 latency — the response time for the slowest 5% of requests — by at least 25% across its Realtime voice models, driven by improved caching. For voice agents, where even brief pauses break the feel of a live conversation, that reduction is the main selling point of the release.
What changed
Alongside the latency work, OpenAI says GPT-Realtime-2.1 brings better alphanumeric recognition — reading out phone numbers, order codes, and similar strings more reliably — plus improved handling of silence, background noise, and mid-sentence interruptions. The model supports configurable reasoning effort, instruction following, and tool use, letting developers trade off latency against answer quality for a given task.
A second, smaller model, GPT-Realtime-2.1-mini, ships alongside it as a distilled version built for faster, lower-cost voice interactions where raw capability matters less than speed and price.
Pricing
According to OpenAI’s model documentation, GPT-Realtime-2.1 is priced at $32 per million audio input tokens and $64 per million audio output tokens (plus $4 and $24 for text input and output). The mini variant costs roughly a fifth as much: $10 per million audio input tokens and $20 per million audio output tokens. Both models carry a 128,000-token context window.
Availability
Both models are live now on the v1/realtime endpoint and can be tried directly in the Realtime Playground on OpenAI’s developer platform.