gpt-realtime
2025-08-28 · Publisher : OpenAI · Domain : Artificial Intelligence
OpenAI makes the Realtime API generally available and launches gpt-realtime, its most advanced speech-to-speech model, with better instruction following, more accurate tool calls and more natural voices.
What changes
- Realtime API in general availability, with remote MCP servers, image input and phone calls via SIP.
- A single model that processes and generates audio directly, without chaining transcription, LLM and speech synthesis.
- Two new voices, Cedar and Marin, exclusive to the Realtime API, and eight existing voices improved.
- Can follow fine-grained instructions ("speak quickly and in a professional tone"), switch language mid-sentence and capture alphanumeric sequences.
- Asynchronous function calls: the conversation continues while a long call is running.
### Specifications - Context 32,000 tokens, maximum output 4,096 tokens, text, audio and image inputs, text and audio outputs. - Knowledge cutoff: October 1, 2023. - API identifier: gpt-realtime (snapshot gpt-realtime-2025-08-28), via WebRTC, WebSocket and SIP. ### Pricing Price per million tokens (USD), down 20% from gpt-4o-realtime-preview. | | Input | Cached input | Output | | --- | --- | --- | --- | | Audio | 32 | 0.40 | 64 | | Text (documentation) | 4 | 0.40 | 16 | ### Benchmarks Scores published by the vendor, against the December 2024 model. | Benchmark | gpt-realtime | December 2024 model | | --- | --- | --- | | Big Bench Audio (reasoning) | 82.8% | 65.6% | | MultiChallenge audio (instruction following) | 30.5% | 20.6% | | ComplexFuncBench audio (function calling) | 66.5% | 49.7% | ### Key takeaways gpt-realtime makes voice agents usable in production: call centers, assistants and education. The MCP and SIP connectivity simplifies integration with existing tools and telephony.
Links
Tags : Audio, LLM, OpenAI, Temps réel, Voix, MCP