All models
Voice & realtimeStable
OpenAI
GPT-Realtime-2.1
OpenAI's realtime speech-to-speech reasoning model for tool-using voice agents that also need text and image context.
Plain-English overview
What GPT-Realtime-2.1 actually is
GPT-Realtime-2.1 is designed for live conversations where the model must listen, speak, follow instructions, and call tools in one realtime session. OpenAI highlights improved interruption, silence, noise, and alphanumeric handling over the prior version.
Audio, text, and image input are billed in different token classes. A realistic budget therefore needs captured usage from full conversations, not a simple conversion from audio minutes alone.
Good fit for
- Tool-using customer or employee voice agents
- Multimodal realtime assistants
- OpenAI-native agent stacks
Category comparison
The facts that matter for voice models
These are provider-published specifications, not Cody benchmark scores. Follow the linked sources for current limits and endpoint-specific exceptions.
- Price basisToken-, character-, or minute-based price for the listed access route.
- Audio $32 in · $64 out / 1M tokens
- Interaction modeSpeech-to-speech, audio-to-audio, or streamed text-to-speech behavior.
- Speech-to-speech with text/image context
- Published latencyProvider-published measurement with its stated exclusions; not a Cody test.
- Not published
- LanguagesProvider-documented language coverage or a clear unpublished marker.
- Multilingual; exact list not published here
- Tools & controlNative function calling, reasoning, or speech-control features.
- Function calling and configurable reasoning
- Context or input limitDocumented token or character limit where it is meaningful.
- 128K input · 32K max output
API and provider access
Where to get GPT-Realtime-2.1
Availability
Regions and access stage
Available through OpenAI's Realtime API in supported countries and territories.
Direct access follows OpenAI's live supported-country list.
Check live availabilityData and training
The route matters.
OpenAI says API content is not used for training by default. Default abuse-monitoring logs may be retained up to 30 days, with additional controls available to qualifying organizations.
This is a concise reading of the cited provider material, not legal advice. A third-party gateway can have different storage, routing, training, and residency terms from the model maker's direct API.
Read the provider policyFrequently asked questions
GPT-Realtime-2.1 FAQ
What is GPT-Realtime-2.1?
OpenAI's realtime speech-to-speech reasoning model for tool-using voice agents that also need text and image context. GPT-Realtime-2.1 is designed for live conversations where the model must listen, speak, follow instructions, and call tools in one realtime session. OpenAI highlights improved interruption, silence, noise, and alphanumeric handling over the prior version.
When was GPT-Realtime-2.1 released?
The provider does not publish a clear release date for GPT-Realtime-2.1 in the source material reviewed by Cody.
Where can I access GPT-Realtime-2.1?
Available through OpenAI's Realtime API in supported countries and territories. The access routes listed in this guide are OpenAI.
How much does GPT-Realtime-2.1 cost?
$32 audio input · $64 audio output / 1M tokens. Text is $4 input and $24 output per million tokens; image input is $5 per million tokens. Cached-input rates differ.
Where is GPT-Realtime-2.1 available?
Available through OpenAI's Realtime API in supported countries and territories. Direct access follows OpenAI's live supported-country list.
Is my GPT-Realtime-2.1 API data used for training?
OpenAI says API content is not used for training by default. Default abuse-monitoring logs may be retained up to 30 days, with additional controls available to qualifying organizations. The policy belongs to the provider route and account terms, so verify it again before production use.
Research trail
Sources for this guide
Keep comparing
More voice & realtime models
Google · Preview
Gemini 3.1 Flash Live Preview
Google's preview audio-to-audio model for low-latency dialogue with multimodal awareness, thinking, search grounding, and function calling.
Read guideElevenLabs · Stable
Eleven v3 Conversational
ElevenLabs' expressive realtime text-to-speech model for natural dialogue, emotional delivery, audio tags, and more than 70 languages.
Read guideInworld AI · Stable
Inworld Realtime TTS-2
Inworld's newest expressive realtime text-to-speech model, designed to remember conversational delivery and speak in more than 100 languages.
Read guide