OpenAI launches GPT-Live for full-duplex ChatGPT Voice
OpenAI launched GPT-Live across paid ChatGPT Voice plans, with a free rollout in progress. The model can listen while speaking and delegate search or deeper reasoning asynchronously.
TL;DR
- GPT-Live is the new ChatGPT Voice stack for paid users, with Go, Plus, and Pro fully rolled out and Free rollout still in progress, per OpenAI's rollout update.
- The architecture is full-duplex: OpenAI's architecture clip says it can listen and speak at the same time, and OpenAI's delegation clip says harder work can run on a frontier model in the background.
- OpenAI is turning voice intelligence into a setting, with TestingCatalog's mobile screenshots showing Instant, Medium, and High options inside the Voice menu.
- Developers do not have day-one API access; OpenAIDevs says GPT-Live-1 and GPT-Live-1 mini are coming to the API soon.
- Day-one limits are real: btibor91's release-note screenshot says GPT-Live does not support video or screen sharing and is not available in Business, Enterprise, or Edu workspaces at launch.
The OpenAI launch post buries the most useful plumbing in the middle: GPT-Live can keep the conversation alive while a heavier model handles search or reasoning. Simon Willison's Weblog quotes OpenAI saying GPT-Live uses GPT-5.5 in the background at launch and will swap in newer frontier models over time. There is also a system card, an API notification form, and a release-note screenshot with caveats the launch thread compresses into one sentence.
What shipped
- Paid ChatGPT Voice is already on GPT-Live-1; OpenAI's rollout update says Go, Plus, and Pro are fully rolled out, with Free rollout in progress.
- Free users get GPT-Live-1 mini by default, according to btibor91's launch notes.
- OpenAI says ChatGPT support spans iOS, Android, and web in its availability post.
- The Voice settings now expose an Intelligence selector with Instant, Medium, and High, shown in TestingCatalog's screenshots and TestingCatalog's selector reply.
- GPT-Live-1 and GPT-Live-1 mini are coming to the API, with OpenAIDevs pointing developers and enterprises to a notification form.
- GPT-Live supports memory and custom instructions, according to juberti's reply, while juberti's later reply says OpenAI is looking into reported memory-access issues.
Full-duplex architecture
Taboo demo with simultaneous speaking and listening
OpenAI's core claim is simple: GPT-Live can listen and speak at the same time. That unlocks fast interruptions, backchannels like “mhmm,” and silence that does not force the model to end a turn.
The mechanics OpenAI calls out:
- continuous audio processing while generating speech, from OpenAI's full-duplex demo
- better back-and-forth timing and live translation, from OpenAI's timing and translation clip
- a more defined conversational personality than earlier voice models, from OpenAI's personality clip
- promptable behavior, according to juberti's reply
- user-controllable speaking speed, according to juberti's reply
The Taboo demo is the cleanest product test in the evidence: GPT-Live plays as a teammate, gets interrupted, and keeps pace without the old walkie-talkie rhythm.
Async delegation
GPT-Live is also a router. It handles the live conversation, then hands off search, deeper reasoning, or more agentic work to a frontier model without freezing the voice session.
Simon Willison's Weblog quotes OpenAI saying GPT-Live uses GPT-5.5 in the background at launch. That same quote says OpenAI will continuously update the delegated model as new frontier models ship.
The useful mental model is a real-time conversation layer over heavier systems. LiorOnAI's breakdown says the system splits timing, reasoning, search, memory, planning, and speech into coordinated parts that still feel like one voice to the user.
Benchmarks
OpenAI's charts compare GPT-Live against AVM, the older Advanced Voice Mode baseline.
From those charts:
- GPQA: AVM 45.3% to GPT-Live-1 High 84.2%, +38.9 points.
- GPQA: AVM 45.3% to GPT-Live-1 mini 74.9%, +29.6 points.
- BrowseComp: AVM 0.7% to GPT-Live-1 High 75.2%, +74.5 points.
- BrowseComp: AVM 0.7% to GPT-Live-1 mini 31.6%, +30.9 points.
- T3-Voice Telecom: AVM sits near 30% success at roughly 380 seconds, while GPT-Live-1 High sits near 63% at roughly the same duration.
- T3-Voice Telecom: GPT-Live-1 Instant trades success for speed, landing near 37% success at roughly 230 seconds.
The benchmark shape matches the product UI: Instant, Medium, and High are not just labels. They expose the latency-quality trade inside a live voice session.
Day-one gaps
The release notes carry the caveats engineers will hit before the demos do.
- GPT-Live does not support video or screen sharing at launch, according to btibor91's release-note screenshot.
- Eligible subscribers can keep using Advanced Voice Mode for those capabilities, according to btibor91's release-note screenshot.
- GPT-Live is not available in Business, Enterprise, or Edu workspaces at launch, according to btibor91's release-note screenshot.
- btibor91's launch notes also say Temporary Chats, the desktop app, Work, Codex, and custom GPTs are not initial surfaces.
- Memory support exists, but juberti's reply says OpenAI is investigating reports from users whose GPT-Live sessions could not access memories.
The most awkward split is Codex. gdb's post says OpenAI is working on bringing GPT-Live to API and Codex, while the day-one ChatGPT rollout keeps Codex outside the launch surface.
Vibe Check
Introducing GPT‑Live
Introducing GPT‑Live OpenAI finally upgraded the model used by ChatGPT voice mode! I've had preview access for a few weeks in the iPhone app, and the new model is very impressive. It also has the ability to spin off harder tasks to GPT-5.5: For questions that require web search, deeper reasoning, or more complex work, it delegates to our latest frontier model behind the scenes and brings the result back into the conversation when it’s ready. While it works, GPT‑Live can keep talking with you and maintain the flow of conversation. At launch, GPT‑Live will use GPT‑5.5 in the background. As we release new frontier models, we’ll continuously update the model used by GPT‑Live. The previous voice mode in the ChatGPT app was based on a GPT-4o era model, with a knowledge cut-off some time in 2024. I had mostly stopped using voice mode because the age and relative weakness of the model greatly limited how useful it was as a brainstorming partner. During the preview period I encountered a pretty obscure bug: the model was interrupting me to laugh at things I said, which weren't even intended as jokes! It felt rude and condescending - I reported it to OpenAI and as far as I can tell they made some tweaks and it's now less likely to happen. From looking back at my transcripts I think it was this bit that triggered the interrupting laugh: so where are the owls when they're not, like before dusk? The owls exist, right? Are they hiding in holes? Where are they hiding? My longest conversatio
Hands-on reports clustered around the same practical change: the model feels useful as a live thinking partner because the voice layer is no longer stuck on an older, weaker model.
- Simon Willison's Weblog says the previous ChatGPT voice mode was based on a GPT-4o-era model with a 2024 knowledge cutoff, which made it much less useful for brainstorming.
- Willison also hit a preview bug where the model interrupted him with laughter, then reported it to OpenAI and saw the behavior become less likely, according to Simon Willison's Weblog.
- Multilingual use looked strong in cedric_chee's hands-on note, which mentions Thai, Japanese, Chinese dialects, Singlish, and live translation that “barely missed a beat.”
- embirico framed the step up around interruption handling, after years of teaching people how to avoid cutting off older voice models.
- Delegation was the standout behavior for jxnlco, who called GPT-Live “great at delegation.”
- sama said he had always preferred typing to talking to AI, but expected that to shift after GPT-Live.
Pre-launch and API context
GPT-Live had already surfaced under another name. TestingCatalog's post said “GPT Bidi 1” had been renamed to “GPT Live 1,” and Kolt Regaskes' post described the same rename as a shift from bidirectional realtime voice mode to GPT-Live.
The API side moved separately before the GPT-Live launch. OpenAIDevs said GPT-Realtime-2.1-mini brought reasoning and tool use to the Realtime mini lineup at the same cost as GPT-Realtime-mini, with p95 latency reduced by at least 25% across Realtime voice models.
The community was already asking for bidirectional voice on that API thread. TestingCatalog's reply answered the Realtime update with “we need Bidi,” and TestingCatalog's Playground screenshot showed GPT-Realtime-2.1 and 2.1-mini available before GPT-Live API access opened.