← All models

GPT-Realtime-2

Published Publisher OpenAIgpt-realtime-2

Overview

GPT-Realtime-2 is OpenAI’s real-time speech-to-speech model for voice agents. OpenAI describes it as its first voice model with GPT-5-class reasoning. It was released on May 7, 2026, and is available through the Realtime API. (openai.com)

Capabilities and uses

  • Supports text and audio input and output, plus image input. It can follow instructions, handle interruptions and corrections, and use functions or parallel tool calls during a conversation.
  • Developers can configure reasoning effort from minimal through xhigh; the announced default is low. Higher effort can increase latency and output-token use.
  • Intended for applications such as customer-support agents, travel assistance, spoken guidance and voice-driven task completion. (openai.com)

Limits

  • Context window: 128,000 tokens; maximum output: 32,000 tokens. Knowledge cutoff: September 30, 2024.
  • Video input, image output, structured outputs and fine-tuning are not supported. Usage limits depend on account tier; the model page lists the free tier as unsupported.
  • Realtime sessions are subject to safety safeguards and may be interrupted when violations are detected. (developers.openai.com)

OpenAI API pricing

The official model page lists the following USD prices per 1 million tokens, checked November 16, 2026: (developers.openai.com)

Modality Input Cached input Output
Text $4.00 $0.40 $24.00
Audio $32.00 $0.40 $64.00
Image $5.00 $0.50 Not supported

Eligible regional-processing endpoints for models released on or after March 5, 2026 incur a 10% uplift. Tool-specific charges may also apply. These are OpenAI API prices; other platforms bill separately. (platform.openai.com)

Official sources