GPT-Realtime-2 - OpenAI's second-generation real-time voice model

What is GPT-Realtime-2

GPT-Realtime-2 is OpenAI's second-generation real-time voice model and is currently the voice Agent with the strongest reasoning capabilities in the Realtime API. The model features GPT-5 -level reasoning capabilities, enabling it to listen and think simultaneously during voice conversations, handle complex multi-step tasks in real time, and precisely invoke external tools. GPT-Realtime-2 supports natural interruptions, topic switching, and error recovery. Its context window has been expanded from 32K to 128K, and it adds a preamble voice feedback feature, allowing voice Agents to truly evolve from being able to chat to being able to get work done.

Main features of GPT-Realtime-2

• GPT-5-level real-time reasoning: handles complex multi-step reasoning tasks during voice conversations, no longer limited to simple Q&A.
• Parallel tool calling: can invoke multiple external tools simultaneously (such as checking calendars, calling APIs, updating CRMs), with real-time voice feedback on execution progress.
• Preamble voice feedback: proactively says "let me check" when performing operations, avoiding users waiting in silence and improving interaction transparency.
• Natural conversation handling: supports interruptions, topic switching, error correction, and context recovery, making the conversation experience feel closer to talking with a real person.
• Five adjustable reasoning effort levels: offers five levels: minimal, low, medium, high, and xhigh, with low as the default, balancing latency and deep thinking.
• 128K context window: four times larger than the previous generation's 32K, supporting longer and more coherent Agent workflows.
• Emotion and tone control: can adjust tone based on the scenario, such as calmly solving problems, empathetically reassuring, or positively confirming.

Technical principles of GPT-Realtime-2

• End-to-end audio understanding: based on the GPT-5 architecture, it maps raw audio directly to semantic representations, skipping the traditional ASR text conversion step.
• Streaming audio token processing: uses streaming audio token encoding to achieve a millisecond-level low-latency full pipeline of "listening-understanding-reasoning-speech synthesis."
• Unified multimodal space: completes audio input, semantic reasoning, tool decision-making, and voice output within a single model, reducing information loss between modules.
• Parallel tool calling engine: while the voice stream continues transmitting, it asynchronously invokes external APIs in the background and converts progress into voice feedback through the preamble mechanism.
• Adjustable reasoning effort control: dynamically balances response speed and reasoning depth through five levels of compute budget allocation from minimal to xhigh.
• Long-range context caching: the 128K context window works with a streaming cache mechanism to support coreference resolution and topic tracking in ultra-long sessions.
• Agents SDK guardrail integration: deeply integrated with OpenAI's safety framework, detecting harmful content in real time and supporting developers' custom business rules.

How to use GPT-Realtime-2

• Obtain access: register for an OpenAI developer account to obtain an API Key, and ensure that Realtime API access has been enabled.
• Choose a protocol: choose WebRTC (low latency in browsers), WebSocket (flexible control), or SIP (telephone systems) access according to the application scenario.
• Create a session: initiate a session request to the Realtime API, specifying the model as gpt-realtime-2, configure audio input/output formats.
• Set reasoning level: choose reasoning intensity based on task complexity (minimal / low / medium / high / xhigh, default low).
• Configure tools: define callable tools via the Agents SDK (such as checking calendars, calling databases, updating CRM), and enable preamble voice feedback.
• Establish audio stream: the client captures the microphone audio stream and sends it to the API, while receiving the speech stream returned by the model.
• Handle interaction: the model reasons while listening, automatically calls tools and reports progress via voice; developers only need to handle business logic and exception fallbacks.

Key information and usage requirements for GPT-Realtime-2

• Product name：GPT-Realtime-2
• Development team：OpenAI
• Integration method：Realtime API（WebRTC / WebSocket / SIP）
• Audio pricing: input $32 / 1M tokens (cached $0.40), output $64 / 1M tokens
• Text pricing: input $4 / 1M tokens, output $16 / 1M tokens
• Context window：128K
• Reasoning levels: minimal / low / medium / high / xhigh (default low)
• Usage requirements: requires an OpenAI API Key, supports quick integration into existing applications via Codex .

Core advantages of GPT-Realtime-2

• Strongest reasoning capability: Big Bench Audio score is 15.2% higher than GPT-Realtime-1.5, with a substantial increase in the completion rate of complex voice tasks.
• Most reliable tool calling: Zillow testing shows that after prompt optimization, the call success rate increased from 69% to 95%, with stronger compliance guardrails.
• Longest context: the 128K window supports Agentic workflows such as ultra-long meetings and in-depth customer service.
• Highest controllability: five reasoning intensity levels + adjustable tone; developers can fine-tune parameters according to business scenarios.
• Most complete ecosystem: seamless integration with OpenAI's full-stack toolchain (Agents SDK, Codex).

Project address of GPT-Realtime-2

• Project official website：https://openai.com/index/advancing-voice-intelligence-with-new-models-in-the-api/

Comparison of GPT-Realtime-2 with similar competitors

Application scenarios of GPT-Realtime-2

• Intelligent customer service: handles complex voice services requiring multi-step operations such as returns and exchanges, appointments, and inquiries, directly calling backend systems to complete the loop.
• Sales Assistant: Retrieve customer information in real time, update CRM, schedule meetings, and report execution results via voice.
• Personal Voice Butler: Manage schedules, order food, and navigate through natural conversation—complete life tasks just by speaking.
• Medical Voice Assistant: Accurately understand professional terminology and symptom descriptions, and provide compliant preliminary consultation and appointment guidance.
• Financial Voice Service: Handle scenarios such as account inquiries, transaction confirmations, and investment advice that require rigorous reasoning and compliance guardrails.
• Enterprise Office Agent: Record in real time during meetings, assign tasks, and access document systems, becoming a true voice office entry point.