Voice-to-action agents
Let callers make requests that require reasoning, API lookups, scheduling, account actions, or multi-step tool use.
Independent tool overview
GPT-Realtime-2 at a glance
GPT-Realtime-2 is OpenAI's reasoning-capable speech-to-speech model for live voice agents, with adjustable reasoning, function calling, interruption handling, expressive delivery, and a 128K context window. It remains available, but GPT-Realtime-2.1 is now the improved same-price update.
Visit the official GPT-Realtime-2 site ↗
- Model ID
- gpt-realtime-2
- Inputs
- Audio, text, and images
- Outputs
- Audio and text
- Context window
- 128,000 tokens
- Max output
- 32,000 tokens
- Current successor
- GPT-Realtime-2.1
Overview
What GPT-Realtime-2 is
GPT-Realtime-2 lets developers build agents that listen, reason, speak, and call tools during a live conversation. It is aimed at customer support, scheduling, commerce, guidance, and other workflows where a voice interface needs to complete tasks instead of merely answering questions.
The model can give a short spoken preamble before a tool call, run multiple tools in parallel, explain what it is checking, and recover verbally when a tool or request fails. Adjustable reasoning levels let builders trade latency and cost for stronger performance on harder turns.
GPT-Realtime-2 is still listed as an active API model. For new deployments, GPT-Realtime-2.1 is the better default because OpenAI says it improves alphanumeric recognition, silence and noise handling, and interruption behavior at the same token prices.
Use cases
Who GPT-Realtime-2 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Customer support calls
Handle natural corrections, interruptions, specialized vocabulary, and empathetic or reassuring delivery.
Live spoken guidance
Turn changing system context into immediate explanations, alerts, and next-step instructions.
Longer voice workflows
Maintain more conversation and tool context across complex sessions with the 128K context window.
Capabilities
Core GPT-Realtime-2 features
Configurable reasoning
Choose minimal, low, medium, high, or xhigh reasoning effort; higher settings can improve hard tasks while adding latency and token use.
Parallel tool calls
Call multiple functions at once when a request needs several independent lookups or actions.
Spoken preambles
Tell the user what the agent is doing with short phrases such as checking a calendar or looking up an order.
Interruption and correction handling
Keep the exchange coherent when a user changes the request or speaks over the response.
Expressive delivery
Adjust tone and speaking style to the situation, from calm troubleshooting to upbeat confirmation.
Image input
Accept still images as additional context alongside live audio and text.
Realtime transports
Connect browser, server, or telephony experiences through WebRTC, WebSocket, or SIP-based Realtime flows.
Process
How the GPT-Realtime-2 workflow works
Step 1
Define the call boundary
Specify what the agent may answer, which actions it may take, and which cases require a human or explicit confirmation.
Step 2
Choose a connection
Use WebRTC for client-side voice, WebSocket for server-side control, or SIP for phone-call integration.
Step 3
Configure voice and reasoning
Start with low reasoning effort, then raise it only for turns where better planning justifies added latency and cost.
Step 4
Add narrow tools
Expose well-described functions with validated parameters, clear success states, and safe failure messages.
Step 5
Test real conversations
Evaluate noise, silence, interruptions, names, numbers, accents, tool failures, latency, disclosure, and escalation paths before launch.
Cost
GPT-Realtime-2 pricing and free plan
GPT-Realtime-2 is billed by text, audio, and image tokens. There is no free API tier. GPT-Realtime-2.1 currently uses the same rates and is the stronger choice for new builds.
Text tokens
$4 input / $24 output per 1M tokens
Text context and generated text or reasoning usage.
- Cached text input is $0.40 per 1M tokens.
- Higher reasoning effort can increase output-token use.
Audio tokens
$32 input / $64 output per 1M tokens
Live spoken input and generated speech.
- Cached audio input is $0.40 per 1M tokens.
- Actual call cost depends on turn length, interruptions, context, and tool behavior.
Image input
$5 per 1M tokens
Still images supplied as context to the voice agent.
- Cached image input is $0.50 per 1M tokens.
- The model does not output images or video.
Pricing checked . Check current pricing at the source ↗
Assessment
GPT-Realtime-2 strengths and limitations
Where it stands out
- Combines low-latency speech with reasoning and reliable function calling.
- Preambles and audible tool transparency make pauses feel less confusing.
- A 128K context window supports longer, more agentic sessions.
- Adjustable reasoning provides a practical latency-versus-intelligence control.
- Supports browser, server, and telephony deployment patterns.
What to consider
- GPT-Realtime-2.1 now improves several production voice behaviors at the same price, so GPT-Realtime-2 is no longer the best starting model.
- Audio output is relatively expensive, and long conversations can accumulate substantial token cost.
- Higher reasoning levels can increase latency, which may make a live call feel less natural.
- Structured outputs, fine-tuning, video, and image output are not supported.
- Developers remain responsible for AI disclosure, consent, recording laws, tool permissions, escalation, and domain-specific compliance.
Compare
GPT-Realtime-2 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
Gemini 3.1 Flash Live
For Google's low-latency live voice model and Gemini ecosystem.
Explore Gemini 3.1 Flash Live →Sales
Vapi
For a managed voice-agent platform that handles telephony and lets teams choose among model providers.
Explore Vapi →Business Operations
Retell AI
For a packaged conversational voice platform focused on production phone agents.
Explore Retell AI →Miscellaneous
Realtime TTS-2
For realtime speech synthesis that adapts delivery to conversational audio rather than running the full agent brain.
Explore Realtime TTS-2 →Questions
GPT-Realtime-2 FAQs
What is GPT-Realtime-2?
It is an OpenAI API model for live speech-to-speech agents that can reason, call functions, accept corrections, and respond with expressive audio.
Is GPT-Realtime-2 still available?
Yes. OpenAI still lists it as an active model, but GPT-Realtime-2.1 is now the improved version and uses the same token prices.
What changed in GPT-Realtime-2.1?
OpenAI says version 2.1 improves alphanumeric recognition, silence and noise handling, and interruption behavior.
How much does GPT-Realtime-2 cost?
Current standard rates are $32 per million audio input tokens, $64 per million audio output tokens, $4 per million text input tokens, and $24 per million text output tokens.
Can GPT-Realtime-2 call tools?
Yes. It supports function calling and can make parallel tool calls during a conversation.
Can it connect to phone calls?
Yes. Realtime applications can use SIP for telephony, while WebRTC and WebSocket cover client-side and server-side connections.
Bottom line
Our GPT-Realtime-2 verdict
GPT-Realtime-2 was a major step from fast voice response to genuinely agentic voice interaction. Its reasoning controls, tool use, recovery behavior, and long context still make it viable, but teams starting now should normally select GPT-Realtime-2.1 and evaluate the full call experience—not just model accuracy—across latency, interruptions, tool failures, safety, and cost.
Visit GPT-Realtime-2 website ↗Similar Miscellaneous Tools
View all Miscellaneous AI tools →
xAI Custom Voices
Custom Voices - xAI's new tool to clone your voice from short clips for use within Grok's applications
AutoScientist
Autoscientist - Adaption's new tool for automating AI model training
Deep Max
Deep Max - Exa's new SOTA agentic search tool
Miso One
Miso One - Open-source text-to-speech model that reads a speaker’s tone for expressive responses
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.