OpenAI has officially released the Realtime API to all developers and enterprises, unveiling a host of new features designed to help teams build production-grade voice agents with greater reliability and flexibility. The API now offers remote MCP server support, seamless image input handling, and phone calling capabilities via Session Initiation Protocol (SIP), unlocking significant new tools and context for AI-powered voice experiences.

Introducing GPT-Realtime: Natural and Expressive Speech

The highlight of the update is the debut of GPT-Realtime, OpenAI’s most advanced speech-to-speech model yet. This model excels in handling complex instructions, precise tool calling, and generating speech that sounds convincingly human—complete with nuanced intonation, emotional cadence, and multilingual proficiency, including mid-sentence language switching. International benchmarks underscore its performance, with GPT-Realtime scoring 82.8% on Big Bench Audio for reasoning tasks and 30.5% on MultiChallenge Audio for instruction-following, leaving previous models in the dust.

Expanded Voice Library: Meet Cedar and Marin

In addition to upgrading audio fidelity across its eight existing voices, OpenAI introduces two new exclusive voices—Cedar and Marin. These voices set a new standard for speech quality, offering more natural, expressive output and smoother conversational experiences for diverse applications, from customer service to education.

Streamlined Architecture: Fast, Low-Latency Deployment

Unlike conventional systems that chain speech-to-text and text-to-speech modules, the Realtime API processes audio directly through a single, unified model. This design innovation minimizes latency, maintains subtle inflections, and bolsters user engagement with fluid, interactive dialogue.

Powerful New Integrations & Capabilities

• Remote MCP Server Support: Easily add new agent capabilities by configuring the API to connect to different MCP servers—no manual wiring required.

• Image Inputs: Users can now submit photos and screenshots alongside audio and text, enabling the model to provide contextually relevant insights and answers based on visual information.

• SIP Phone Calling: The API now natively connects to PBX systems, desk phones, and public phone networks, further expanding use cases in telecommunications and support.

• Reusable Prompts: Developers can save rich, programmable prompts—complete with context and example messages—for future use, driving consistency and flexibility across sessions.

Safety, Privacy, and Compliance

OpenAI has incorporated robust classifiers and multi-layered safeguards to prevent misuse and ensure responsible AI deployment. The Realtime API supports EU Data Residency requirements and adheres to stringent enterprise privacy commitments, giving organizations control over sensitive data. Usage policies strictly prohibit malicious use, with clear guidelines for transparency during AI interactions.

Pricing and Availability

Starting today, the Realtime API and GPT-Realtime model are open to all developers, with pricing reduced by 20% compared to previous versions—$32 per million audio input tokens and $64 per million audio output tokens. Cached tokens and advanced session management tools further reduce costs and allow for scalable deployment across prolonged conversations.

For detailed pricing, please visit the official detailed pricing page.

Livestream