Faster voice AI: Realtime 2.1 cuts p95 latency by at least 25% through improved caching.
More capable agents: The API supports audio, images, function calls, MCP, WebRTC, WebSocket, and SIP.
More pricing flexibility: Realtime 2.1 mini offers lower audio costs for high-volume voice applications.
OpenAI Realtime API gives software a direct path to natural voice conversations. A traditional voice system may turn speech into text, send it to a language model, then turn the answer back into speech. Realtime can handle speech-to-speech dialogue in one model flow. This design cuts delay and helps an AI agent react to pauses, interruptions, noise and changes in speech.
The latest models, GPT-Realtime-2.1 and GPT-Realtime-2.1 mini, arrived in July 2026. OpenAI says these models cut p95 latency by at least 25% across Realtime voice models through better cache use. GPT-Realtime-2.1 also adds better alphanumeric recognition, silence and noise control, interruption behavior, instruction control and tool use. These changes matter when a wrong number or poor interruption response can harm a call.
A Realtime session keeps a live connection between an app and the model. Audio can move into the session and back from the model with very little delay. The API supports WebRTC, WebSocket and SIP, so developers can choose a setup that fits the product.
WebRTC fits browser and mobile apps. WebSocket suits server-side systems. SIP opens a path to phone systems and call centers. The model can also use text and images along with audio. Realtime 2.1 supports function calls, so the model can ask an external system for an order check or account lookup.
This setup can turn a voice assistant into a useful agent rather than a simple chatbot. A customer can ask about an order, and the agent can check company data before it gives an answer. The same process can work with account details, appointments, support cases and other business tools.
GPT-Realtime-2.1 targets cases that need better logic, tool use and instruction control. Its price is $4 per 1 million text input tokens, $24 per 1 million text output tokens, $32 per 1 million audio input tokens and $64 per 1 million audio output tokens. Image input costs $5 per 1 million tokens; cached audio input costs $0.40 per 1 million tokens.
GPT-Realtime-2.1 mini targets high-volume and cost-sensitive voice products. It costs $0.60 per 1 million text input tokens, $2.40 per 1 million text output tokens, $10 per 1 million audio input tokens and $20 per 1 million audio output tokens. Image input costs $0.80 per 1 million tokens; cached audio input costs $0.30 per 1 million tokens. Both models have a 128K context window.
The price gap matters at large call volumes. Realtime 2.1 mini has audio input and output rates about 3.2 times lower than the full 2.1 model. This makes the mini model a strong fit for services with many simple voice requests.
Also Read - OpenAI Responses API: Features, Use Cases, Examples
The API fits products where voice forms the core experience. Customer support shows the value. A caller can explain a problem while the agent checks account data, finds an order, calls a tool and gives an answer.
Phone agents offer another strong use case. SIP support lets a voice agent connect to phone infrastructure rather than old menu trees. The agent can choose the right tool or service from plain speech.
The API also fits language practice, hands-free assistants and field support. These products need quick replies and a natural flow. A delay can make a voice conversation feel stiff, even when the answer itself remains correct.
Realtime also fits cases where an agent must act during a conversation. A voice assistant can check stock, access a customer record or request a service through an external tool. That ability gives the API a clear role in customer service and other task-based voice products.
Realtime does not suit every AI task. A text chatbot can use the standard Responses API with less system complexity. A large batch of recorded calls may suit a transcription model. A product that only needs speech-to-text does not need a full speech-to-speech agent. A simple text-to-speech feature also needs no Realtime session.
This distinction can save both development effort and API cost. Realtime makes the most sense when a product needs live voice dialogue, quick replies and direct tool access. A standard API can offer a cleaner design when those needs do not exist.
Also Read - 5 Ways OpenAI’s NextSlide Acquisition Could Transform ChatGPT
OpenAI's July 2026 launch of GPT-Live shows the wider direction of voice technology. GPT-Live uses a full-duplex design, so the model can listen and speak at the same time. It can react to pauses, interruptions and changes in pace before each turn ends.
That shift gives the Realtime API a larger role. It no longer looks like a basic voice layer for a chatbot. It supports agents that can hear, reason, speak and act in one live conversation.
For serious voice products, GPT-Realtime-2.1 offers the stronger option, while GPT-Realtime-2.1 mini gives high-volume products a lower-cost path. The right choice depends on call complexity, tool use, response quality and total audio volume.
The OpenAI Realtime API helps developers build live voice applications that can hear speech, respond with audio, understand images, and call external tools.
The latest models are GPT-Realtime-2.1 and GPT-Realtime-2.1 mini. Both offer a 128K context window and support audio, text, images, and function calling.
Realtime 2.1 costs $32 per 1 million audio input tokens and $64 per 1 million audio output tokens. Realtime 2.1 mini costs $10 and $20 respectively.
The API fits live voice agents, customer support, phone systems, language practice, hands-free assistants, and services that need real-time tool use.
Yes. SIP support allows Realtime agents to connect with phone infrastructure, which makes the API useful for call centers, phone support, and conversational IVR systems.