Quick answer
Choose by the job your application needs to do:
- A conversational voice agent: Start with
gpt-realtime-2.1; OpenAI’s Realtime quickstart uses it in a voice-agent example. The model catalog lists other Realtime options too, so compare their current specifications before choosing. - Live speech translation: Shortlist GPT-Realtime-Translate. OpenAI describes a dedicated streaming translation workflow. Confirm the selectable model identifier and session configuration in the translation guide before implementation.
- Live transcription without a spoken assistant reply: Start with
gpt-live-transcribe, which OpenAI’s transcription guide recommends for live audio-to-text. The model catalog also lists GPT-Realtime-Whisper; check its model documentation if you are considering that option.
These are distinct workflows: translation streams translated audio and transcript updates; transcription produces text from live audio; a voice-agent session manages a spoken conversation. A model name alone does not establish that the workflows or their configuration are interchangeable.
OpenAI Realtime models by use case
| If your application needs… | Shortlist | What to check |
|---|---|---|
| A conversational voice agent that can handle turns and use tools | gpt-realtime-2.1; compare other listed Realtime models |
OpenAI’s Realtime guide uses gpt-realtime-2.1 in a voice-agent example. Evaluate the exact model you plan to deploy. |
| Streaming speech translation | GPT-Realtime-Translate | OpenAI describes translation as a dedicated workflow that streams translated audio and transcript updates. Confirm the selectable model ID and session requirements in the translation guide. |
| Streaming transcription or captions | gpt-live-transcribe; also review GPT-Realtime-Whisper |
OpenAI’s transcription guide recommends gpt-live-transcribe for live audio-to-text. Check the relevant guide for the model and configuration you intend to use. |
For voice agents: choose conversation, not just audio input and output
A Realtime voice-agent workflow is for spoken conversations. OpenAI’s current getting-started guide uses gpt-realtime-2.1 in a voice-agent example. Evaluate interruption handling, tool behavior, response quality, and latency with your own prompts and audio rather than relying on a model name alone.
OpenAI’s May 7, 2026 announcement positioned GPT-Realtime-2 for production voice agents, highlighting harder requests, taking action, handling interruptions, and keeping conversations flowing. That is high-level product positioning, not a technical comparison of every model in the current catalog.
For live translation: use the dedicated translation workflow
Choose GPT-Realtime-Translate as a candidate when your task is to translate speech as a person speaks. OpenAI describes streaming translation with translated audio and transcript updates. Its May 2026 announcement reported more than 70 input languages and 13 output languages. These are aggregate counts; confirm the specific languages and directions your application needs in the current translation documentation.
The announcement gives a useful task match, but it does not establish the selectable API identifier. Confirm that identifier and the required session configuration in OpenAI’s Realtime translation guide before using it in production.
For transcription: use transcription mode when you only need text
If you need captions, notes, or a transcript from a microphone or live call—but do not need a spoken assistant response—start with the Realtime transcription workflow. OpenAI’s guide recommends gpt-live-transcribe and describes transcript updates as speech arrives, followed by a final transcript when an audio turn is committed.
The model catalog also lists GPT-Realtime-Whisper for realtime transcription, and OpenAI’s May 2026 announcement described it as streaming speech-to-text. Do not assume the two names are interchangeable: confirm the model ID and configuration in the documentation for the workflow you intend to use.
Production checklist: confirm these before you build
- Exact model ID and availability. Confirm the selectable identifier and that it is available to your project in the documentation for your chosen workflow. Do not treat an announcement’s list as a complete current catalog.
- Session and connection requirements. Check the workflow-specific guide for the correct session type, endpoint, and connection setup. Do not infer translation or transcription setup from a voice-agent example.
- Inputs, outputs, and events. Confirm the required audio format, output modality, transcript events, and any task-critical language coverage in the relevant guide. These details can vary by workflow.
- Credential handling. Follow the current Realtime guide for the chosen transport and authentication flow; do not put a standard API key in browser code.
- Evaluate task-specific behavior. Test the languages, accents, interruptions, terminology, tool calls, and audio conditions that matter to your product. Announcement-level descriptions are useful for shortlisting, not a substitute for model-specific specifications or your own evaluation.
Keep GPT-Live separate from Realtime
OpenAI documents GPT-Live as a voice experience with a separately selected backend for delegated work. Its guide is not a Realtime model catalog, so do not assume GPT-Live setup, connection, or billing details apply to Realtime API sessions. Use the Realtime-specific model and session documentation for a Realtime voice agent.
Bottom line: Shortlist gpt-realtime-2.1 for a conversational voice agent, GPT-Realtime-Translate for live translation, and gpt-live-transcribe for live text transcription. Confirm the selectable identifier, session requirements, language coverage, and current availability in the documentation for that specific workflow before production.
Sources
Getting started with the Realtime API | OpenAI API
Realtime translation | OpenAI API
Realtime transcription | OpenAI API
Advancing voice intelligence with new models in the API | OpenAI