gpt-realtime-2

Direct from Azure models are a select portfolio curated for their market-differentiated capabilities:

• Secure and managed by Microsoft: Purchase and manage models directly through Azure with a single license, consistent support, and no third-party dependencies, backed by Azure's enterprise-grade infrastructure.
• Streamlined operations: Benefit from unified billing, governance, and seamless PTU portability across models hosted on Azure - all part of Microsoft Foundry.
• Future-ready flexibility: Access the latest models as they become available, and easily test, deploy, or switch between them within Microsoft Foundry; reducing integration effort.
• Cost control and optimization: Scale on demand with pay-as-you-go flexibility or reserve PTUs for predictable performance and savings.

Learn more about Direct from Azure models .

About this model

Gpt-realtime-2 is a next‑generation, low‑latency streaming model designed for real‑time speech‑to‑speech and conversational AI scenarios. It processes audio input continuously and generates responses during live interactions, enabling applications such as voice assistants and interactive agents. As part of the latest set of voice models, it builds on existing realtime capabilities to support more advanced conversational experiences within live audio pipelines that combine speech recognition, reasoning, and response generation in a unified flow.

Supported region: Sweden Central, East US 2, Central US, France Central, South India, Canada Central

Key model capabilities

Key Features:

• Speech‑to‑speech (S2S) interaction
  Accepts audio input and produces audio output, enabling full conversational voice flows.
• Built‑in reasoning capabilities
  Incorporates internal reasoning during each turn, improving instruction following and response quality without exposing reasoning traces.
• Reasoning‑first interaction model
  Designed with reasoning‑centric turn behavior, allowing the model to think before producing final responses.
• Configurable reasoning effort
  Supports a reasoning.effort parameter with values:
  minimal | low | medium | high, enabling control over reasoning depth and latency tradeoffs.
• Structured response phases
  Outputs responses in phases:
  commentary → preamble (thinking / filler / tool signals)
  final_answer → final response after reasoning
• Low‑latency streaming for real‑time use
  Optimized for interactive voice applications where responses must be generated quickly during live conversations.
• Long conversational context support
  Supports extended context windows (up to ~128k) to maintain richer multi‑turn conversations.
• Improved instruction following vs earlier models
  Designed to provide stronger adherence to system prompts and user instructions compared to earlier realtime models.
• Successor to gpt‑realtime‑1.5
  Represents a reasoning-enabled upgrade to the previous realtime S2S model stack.