GPT-Realtime-2: OpenAI Quadruples Context Window and Brings GPT-5 Reasoning to Voice AI

A 4× context window expansion, GPT-5-class reasoning, and two companion models for live translation and transcription. With GPT-Realtime-2, OpenAI is not shipping a patch. This is a full repositioning of its voice AI strategy.

GPT-Realtime-2: what the new voice model actually changes

The headline number: context window jumps from 32,000 to 128,000 tokens. For developers building voice agents, that means longer, more complex conversations without memory loss. OpenAI claims an 11% performance gain over GPT-Realtime-1.5, which shipped just three months ago.

On the MultiChallenge audio benchmark, the model scores 30.5% versus 20.6% for its predecessor. ComplexFuncBench, which measures function-calling accuracy, sees a jump to 66.5% from 49.7%. Those are real gains for production voice apps.

Three voice models, not one

Alongside Realtime-2, OpenAI launches GPT-Realtime-Translate for live translation at $0.034 per minute and GPT-Realtime-Whisper for streaming transcription at $0.017 per minute. Two new voices, Cedar and Marin, also join the API roster exclusively.

Pricing on the core model stays flat at $32 per million audio input tokens and $64 per million output tokens. Prompt caching remains the biggest cost lever, cutting input costs by roughly 80×.

GPT-Realtime-2 and the voice AI competition

This release puts OpenAI well ahead in the speech-to-speech API market. Google offers voice capabilities through Gemini, but nothing this dedicated. xAI recently launched voice cloning via API, though it targets a different use case entirely.

The real signal here: voice agents are no longer experimental. Native SIP telephony support, MCP server integration, parallel tool calls, and configurable reasoning effort levels (from minimal to xhigh) all point to one goal. OpenAI wants its model answering your customer service calls, and it wants you unable to tell the difference.

The bottom line on GPT-Realtime-2

The Realtime API officially exits beta and goes GA today. For developers, the message is clear: OpenAI now treats voice as a first-class citizen, right alongside text and code.