Quick answer: GPT-Live is OpenAI’s continuous voice-model family for ChatGPT Voice. Unlike a turn-based voice system, GPT-Live can process incoming speech while it is generating speech, so it can listen through pauses, respond to interruptions, and keep a conversation moving while deeper work runs in the background. OpenAI began rolling out GPT-Live-1 and GPT-Live-1 mini on July 8, 2026.
The practical distinction is not simply “better speech.” GPT-Live separates the live conversation from search, reasoning, and longer-running work. That architecture changes how a user should evaluate voice interaction—and it also creates important product, API, and availability boundaries.
Source check: August 1, 2026. This guide includes OpenAI’s dated July 31 SynthID and content-verification update.
What GPT-Live-1 changes
Earlier voice systems commonly used one of two patterns:
- Cascaded voice: speech is transcribed, a text model answers, and another model converts the answer back to audio.
- Turn-based voice: one multimodal model handles audio, but the user and model still take discrete turns.
OpenAI describes GPT-Live as full duplex. It continuously processes input while generating output and can decide many times per second whether to speak, keep listening, pause, interrupt, or invoke a tool.
That makes several conversation patterns possible:
- Brief acknowledgements without taking over the conversation.
- More natural handling of pauses and interruptions.
- Live translation within the continuous interaction loop.
- Background delegation without ending the active voice exchange.
These are product behaviors described in OpenAI’s launch record, not a guarantee that every language, device, account, or environment will behave identically.
The two-layer architecture
GPT-Live is best understood as two cooperating layers.
| Layer | Primary job | Decision boundary |
|---|---|---|
| GPT-Live-1 | Continuous listening, speaking, timing, and turn management | Best evaluated for conversation flow, interruption handling, and audio behavior |
| Delegated frontier model | Search, deeper reasoning, and more complex work | The model used behind the scenes can change; verify current product documentation |
At launch, OpenAI said GPT-Live used GPT-5.5 in the background. That is a launch fact, not a durable routing promise. OpenAI explicitly said the delegated frontier model would be updated over time.
For current reasoning-model details, use the live OpenAI product documentation. For a source-safe overview of the current GPT family, see the GPT-5.6 model guide.
GPT-Live-1 vs GPT-Live-1 mini
OpenAI introduced two variants:
- GPT-Live-1: announced as the default ChatGPT Voice model for Go, Plus, and Pro users.
- GPT-Live-1 mini: announced as the default ChatGPT Voice model for Free users.
The launch article began a global rollout across iOS, Android, and ChatGPT.com. A rollout statement does not establish access for every account at the same moment, and it does not establish availability in a separate product such as Flowith.
Check the live ChatGPT product for current account access. Check the live Flowith workspace independently before assuming that GPT-Live-1 is a selectable Flowith model.
Is GPT-Live-1 available through the API?
Do not treat the ChatGPT rollout—or the provenance update—as proof that gpt-live-1 is a generally selectable API model ID.
OpenAI’s July 8 launch article said it planned to bring GPT-Live-1 and GPT-Live-1 mini to the API. A dated July 31 update now says supported GPT-Live audio generated through ChatGPT Voice and the OpenAI API includes SynthID. That is a real generation-API claim, but it does not name the callable model ID, access tier, region, pricing, or entitlement.
Current OpenAI Realtime documentation directs developers building low-latency voice agents to gpt-realtime-2.1; it lists separate current models for realtime translation and transcription. Use the live model catalog and developer guide for the exact interface you intend to call rather than substituting the ChatGPT product name.
SynthID and GPT-Live audio provenance
OpenAI’s July 31 update adds a separate question: can a file carry evidence that supported OpenAI tooling generated it?
| Question | Current official answer | What it does not prove |
|---|---|---|
| Does supported GPT-Live audio carry a signal? | OpenAI says supported GPT-Live audio generated through ChatGPT Voice and the OpenAI API now includes SynthID. | It does not mean every historical, transformed, unsupported, or third-party file has a detectable signal. |
| Can a person check a file? | Yes. The public verifier checks supported image and audio files for OpenAI provenance signals. | It is not a general-purpose detector for all AI-generated media. |
| Can a developer automate checks? | Yes. The separate Content Provenance API accepts one file at POST /v1/content_provenance_checks. | Verifier access does not establish access to a particular audio-generation model. |
What does detected mean? | The verifier found a recognized supported signal. | It does not reveal the prompt, account, or individual creator. |
What does not_detected mean? | The verifier did not find a supported signal in that file. | It does not prove human authorship or rule out OpenAI or another AI provider. |
For consequential moderation, journalism, compliance, or trust decisions, inspect the original file when possible and combine the result with source records and human review. Compression, conversion, metadata removal, or other transformations can erase or weaken provenance evidence.
Launch limitations to check again
At launch, OpenAI said GPT-Live in ChatGPT did not support voice with video or screen sharing. Legacy voice modes remained available for those features.
OpenAI also noted that:
- The model was optimized for some of ChatGPT’s most popular languages.
- Some languages could have non-native accents or fluency gaps.
- GPT-Live used predefined ChatGPT voices and was designed for conversation, not voice impersonation.
These details are time-sensitive. Recheck the live product and official documentation before using this article for procurement, accessibility, localization, safety, or implementation decisions.
A practical evaluation checklist
Use a repeatable conversation instead of judging from a short demo:
- Interruption: correct the model halfway through an answer and check whether it recovers without repeating the entire response.
- Pause handling: pause while thinking and see whether the model waits or incorrectly ends your turn.
- Background reasoning: ask a question that needs search or reasoning, then observe whether the conversation remains understandable while work continues.
- Language fit: test the accents, names, and terminology your real users will use.
- Visual handoff: confirm whether a visual answer, file, or image actually appears on the product surface you intend to use.
- Availability: verify the current account, platform, region, and API status directly in official product surfaces.
For the concise model identity and source ledger, use the GPT-Live-1 model page. To separate product choice from model choice, compare Flowith and ChatGPT.
Frequently asked questions
What is GPT-Live?
GPT-Live is OpenAI’s voice-model family for continuous human-AI conversation. GPT-Live-1 and GPT-Live-1 mini were introduced on July 8, 2026 to power a new ChatGPT Voice experience.
What does full duplex mean for voice AI?
It means the system can listen and speak at the same time instead of forcing the user and model into rigid turns. GPT-Live continuously processes input while generating output.
Does GPT-Live-1 replace the reasoning model?
No. OpenAI describes GPT-Live as the live interaction layer. It can delegate search, reasoning, or more complex work to a frontier model in the background.
Can developers call GPT-Live-1 through an API?
Do not infer general GPT-Live-1 model access from the July 31 watermark update. The update establishes that supported GPT-Live audio generated through the OpenAI API includes SynthID, while current Realtime documentation directs low-latency voice-agent builders to gpt-realtime-2.1. Check the live model catalog for the exact model ID.
Does supported GPT-Live audio include a watermark?
OpenAI’s July 31, 2026 update says supported audio generated with GPT-Live through ChatGPT Voice and the OpenAI API now includes SynthID watermarking.
How can I verify GPT-Live audio?
Use OpenAI’s public provenance verifier or the separate Content Provenance API at POST /v1/content_provenance_checks. Both check for supported OpenAI signals; neither is a universal AI detector.
Does not_detected prove that audio is human-made?
No. It only means the verifier did not find a supported signal. A signal can be absent, legacy, stripped, or degraded, and the verifier does not detect every other provider’s AI output.
Is GPT-Live-1 available on Flowith?
This guide does not claim Flowith availability. The official sources establish OpenAI and ChatGPT behavior, not a Flowith model mapping; check the live Flowith workspace independently.
Primary sources
- Introducing GPT-Live — OpenAI’s July 8 launch record and dated July 31 provenance update.
- ChatGPT Voice documentation — current official duplex-model and API-key boundary.
- OpenAI Realtime API guide — current developer source for supported realtime interfaces and models.
- OpenAI provenance verifier — supported signals, public file check, and result-interpretation boundaries.
- Content Provenance API guide — current endpoint, response semantics, limits, and safe-use guidance.