LiveLingoQuick Answer: Can GPT-Live do live translation?
Yes, and it holds the interpreter role reliably, but it falls progressively behind on continuous speech. In ChatGPT, start a Voice session and say something like "simultaneously translate whatever I say into Spanish." Because the models are full-duplex, the translation streams while you are still talking. Free accounts get GPT-Live-1 mini; paid Go, Plus, and Pro plans get GPT-Live-1. Since September 10, 2026 developers can also call it directly as gpt-live-1 at v1/live/sessions for $0.05 per minute plus backend model usage. Measured over a 452-second session, its completion lag climbed from 5.7 seconds to a 10 to 14 second plateau in English to Spanish, and in English to Japanese it translated only 17% of sentences within 30 seconds.
Tap and speak in English
Tap to start
1. What OpenAI Shipped, and What Changed in September
OpenAI announced GPT-Live on Wednesday, July 8, 2026 in a blog post and livestream (Introducing GPT-Live). Two models shipped: GPT-Live-1, the new default voice model for paid Go, Plus, and Pro plans, and GPT-Live-1 mini, the new default for free accounts (per MacRumors and The Decoder). Both replace Advanced Voice Mode as the ChatGPT Voice default, with the old mode remaining selectable in voice settings.
On September 10, 2026 the models reached the OpenAI API, which is the single most important change since launch. GPT-Live-1 is now callable as gpt-live-1 at the v1/live/sessions endpoint, priced at $0.05 per minute of conversation for the voice layer, with whatever backend model handles reasoning billed separately (model reference). Until then the translation claim rested on OpenAI's demo and press reports. It can now be measured directly, and Section 5 does exactly that.
The defining feature is full-duplex audio: the model processes what it hears while it is speaking, making interaction decisions many times per second, so it can be interrupted naturally, respond with backchannels ("mhmm," "got it") while you talk, or stay silent while you think (MarkTechPost). For anything requiring web search, deeper reasoning, or agentic work, GPT-Live delegates to GPT-5.5 in the background and keeps the conversation going while the result comes back. Users can pick three reasoning levels (Instant, Medium, High); the higher levels route to GPT-5.5 Thinking.
OpenAI's published evaluations focus on capability rather than speed: at High reasoning, GPT-Live-1 scores 84.2% on GPQA versus 45.3% for Advanced Voice Mode, and 75.2% on BrowseComp versus 0.7%. In human preference testing, GPT-Live-1 was preferred over Advanced Voice Mode in 75.7% of conversations, and mini in 69.2% (The Decoder, citing OpenAI's announcement). Independently, Artificial Analysis places GPT-Live-1 second overall on its Speech-to-Speech index at 81.5, with 1.34 seconds to first audio. OpenAI still publishes no translation-specific latency figures.
2. How Live Translation Works on GPT-Live
Live translation is one of the capabilities OpenAI explicitly names in the launch materials: the full-duplex architecture lets the model "engage in more natural back-and-forth, maintain a better sense of time, and even perform live translation" (OpenAI announcement). In press briefings, OpenAI demoed the model speaking a running translation while the presenter talked.
It is a prompt, not a mode
There is no translator toggle, language-pair picker, or dedicated translation UI. You start a Voice session and instruct the assistant: "I'd like you to simultaneously translate whatever I'm saying into Hindi." The model then renders a spoken translation as you speak, and continues until you tell it to stop or switch. This is the same prompt-invoked pattern ChatGPT Voice has had since the May 2026 Realtime release; what changed is the full-duplex delivery, which lets the translation overlap your speech instead of waiting for your turn to end.
The role holds, which is not obvious
The obvious failure mode for a conversational model told to interpret is that it answers the speaker instead of translating them. That did not happen. Across 27 minutes of interpreted audio in measured testing, spanning English to Spanish, English to Japanese and Chinese to English, it never once responded to the speaker instead of translating: no clarifying questions, no commentary, no drift back into assistant behavior. Instruction-following is not where this breaks down.
The assistant-first trade-off
GPT-Live is a general conversational assistant that translates on request, not a dedicated interpreter, and that shape has consequences for pace rather than obedience. A conversational model speaks at a natural rate and neither compresses nor drops content when it falls behind. Interpretation has the opposite requirement: the speaker keeps going regardless, so a system that cannot keep up must summarize, skip, or accept unbounded delay. In the consumer app the backchannel behavior that makes conversation feel natural ("mhmm," "yeah") also injects non-translation audio into a translation session, and day-one users widely reported it as intrusive (BigGo).
3. GPT-Live vs OpenAI's Purpose-Built Translator
OpenAI ships two different live-translation surfaces, and they are different stacks. On May 7, 2026, OpenAI released gpt-realtime-translate in the Realtime API: a purpose-built streaming speech-to-speech translation model trained on thousands of hours of professional interpreter audio, configured to remain translation-only and wait for enough context before producing speech (OpenAI announcement). GPT-Live is the general-purpose full-duplex assistant that translates when asked. Both are now available to developers, which makes a direct comparison possible for the first time.
GPT-Live (gpt-live-1) | gpt-realtime-translate | |
|---|---|---|
| What it is | Full-duplex voice assistant; translation by prompt | Dedicated streaming speech-to-speech translation model |
| Released | ChatGPT July 8, 2026; API September 10, 2026 | May 7, 2026 |
| Access | ChatGPT app plus API endpoint /v1/live/sessions | Realtime API, endpoint /v1/realtime/translations |
| Price | Included in ChatGPT plans; $0.05/min via API plus backend model | $0.034 per minute of audio |
| Languages | Unpublished; "most spoken languages," accent gaps disclosed | 70+ input languages; 13 output languages (documented) |
| Stays in translator role | Yes when instructed; zero role breaks in 27 minutes of testing | Yes; translation-only by design, no system prompts |
| Completion lag, en→es | 11.72 s median, rising to a 10–14 s plateau | 2.65 s median, flat |
| Japanese coverage within 30 s | 17% | 98% |
| Transcripts | Source and output transcript deltas | Text deltas of source and translated speech |
The 13 documented output languages of gpt-realtime-translate: English, Spanish, Portuguese, French, German, Italian, Russian, Chinese, Japanese, Korean, Hindi, Indonesian, and Vietnamese (OpenAI Cookbook). OpenAI still publishes no GPT-Live language list, so measured behavior is the only guide, and it varies enormously by language.
4. Limits OpenAI and Independent Testing Disclose
- Lag accumulates on continuous speech. Measured completion lag rose from 5.7 seconds to a 10 to 14 second plateau within three minutes in English to Spanish, and stayed there for the rest of the session.
- Japanese degrades severely. 17% of sentences translated within 30 seconds, against 98% for the same model into Spanish, with 20 of 90 sentences never translated at all.
- Usage caps are real but unpublished. OpenAI's help center states voice usage limits vary by plan and Voice option and may change. Third-party reports of specific numbers conflict; treat any specific figure as unverified.
- No official latency numbers. OpenAI released capability benchmarks (GPQA, BrowseComp, preference rates) but no first-audio or lag measurements for translation.
- Language coverage is undocumented. "Most spoken languages," with non-native accents and fluency gaps disclosed for others.
- Backchannels cannot be fully disabled in the consumer app, though an explicit translation-only instruction held without exception in API testing.
- Safety monitoring can interject. Per the GPT-Live system card, real-time monitoring can steer or interrupt a response, play a spoken safety message, or end high-risk conversations.
5. What Independent Measurement Shows
What we measured (and what we did not)
These numbers come from gpt-live-1 through the OpenAI API, prompted to act as a simultaneous interpreter, measured September 15, 2026. They are not measurements of the ChatGPT consumer experience: free accounts run GPT-Live-1 mini, and the app adds client-side voice detection and its own prompting that no harness can isolate. Every system in the comparison received identical audio at identical real-time pace, and per-clip data is published at livelingo.io/research/gpt-live-1-interpreter-test.
| System | en→es lag | en→ja lag | Japanese coverage within 30 s | Across the session |
|---|---|---|---|---|
| LiveLingo | 0.94 s | 0.88 s | 100% | Flat |
| Gemini 3.5 Live Translate | 1.51 s | 1.92 s | 99% | Flat |
OpenAI gpt-realtime-translate | 2.65 s | 3.33 s | 98% | Flat |
| GPT-Live-1 (prompted) | 11.72 s | n/a | 17% | Rises to 10–14 s |
It never breaks character. Across 27 minutes of interpreted audio it never answered the speaker instead of translating them, in either direction. That is genuinely difficult, and the dedicated translation models do not even have to attempt it.
It falls progressively behind. In English to Spanish the completion lag climbed from 5.7 seconds at minute zero to a plateau between 10 and 14 seconds by minute three, and stayed there. The shape matters: it is not unbounded, but a listener joining at minute six hears a translation roughly twelve seconds behind the room.
English to Japanese collapses. Only 17% of sentences were translated within 30 seconds, against 98% for the same model into Spanish. The median output landed 50.7 seconds late, and 20 of 90 sentences were never translated at all. It stopped emitting about seven seconds after the audio ended rather than draining a backlog, so it did not catch up, it abandoned what it had not reached. The result reproduced across two independent runs.
This is not an OpenAI problem. On the identical Japanese audio, from the same account on the same day, OpenAI's own gpt-realtime-translate held 3.33 seconds flat and covered 98% of sentences. The failure belongs to prompting a full-duplex conversational model to interpret, not to the vendor, and the Japanese figure is why the lag column above reads n/a: when a fifth of the content never arrives, a median describes only the part that survived.
6. How to Use GPT-Live for Translation Today
- Update the ChatGPT app on iOS or Android (or use chatgpt.com). Free accounts run GPT-Live-1 mini; Go, Plus, and Pro accounts run GPT-Live-1.
- Tap the Voice icon in the message composer to start a session.
- Give the translation instruction up front: "Simultaneously translate everything I say into Japanese. Do not answer questions, do not comment, only translate." In measured testing this instruction held without exception.
- For two-way conversation, ask for both directions: "Translate my English into Japanese, and translate any Japanese you hear into English."
- Keep sessions short. The measured lag is lowest in the first minute and roughly doubles by minute three, so breaking a long conversation into shorter exchanges keeps the translation closer to the speaker.
- Developers can call
gpt-live-1directly atv1/live/sessions, or usegpt-realtime-translateinstead if the task is translation rather than conversation.
7. When GPT-Live Fits, and When It Does Not
GPT-Live is the right choice when
- You want free, zero-install live translation and already use ChatGPT. The free tier includes a full-duplex voice model.
- The conversation is short: asking directions, a quick exchange, travel small talk. The drift never has time to accumulate, and the first exchange feels immediate.
- You value the assistant around the translation: you can ask follow-up questions, get context, or have GPT-5.5 look something up mid-conversation.
A different tool fits better when
- The speech is continuous. A meeting, a lecture, or a call where one side talks for minutes at a time is where the measured lag plateau at 10 to 14 seconds becomes the whole experience.
- You are translating into Japanese, or any language where you have not verified behavior. Coverage varied from 98% to 17% between two languages on the same model.
- You need a readable two-language transcript while listening. GPT-Live gives you a chat log, not a side-by-side source-and-translation view that never rewrites itself.
- You need translated phone calls. None of the OpenAI surfaces dial a phone number. LiveLingo places outbound calls to regular phone numbers with translation running on the line, where the recipient needs no app.
An honest concession. LiveLingo (publishing this guide) sits in the dedicated-translator category, so read the comparison with that in mind. GPT-Live's conversational naturalness is a real advance, its full-duplex turn-taking is ahead of every dedicated translator app including ours, its free tier makes it the easiest way for anyone to try live voice translation today, and it held the interpreter role for 27 minutes without a single lapse. What it does not do is keep pace with a speaker who does not wait. Side-by-side specs: /compare/chatgpt-translation. Full measurements: /research/gpt-live-1-interpreter-test.
8. Frequently Asked Questions
What is GPT-Live?
GPT-Live is OpenAI's full-duplex voice model family, debuted in ChatGPT on July 8, 2026. Two models shipped: GPT-Live-1 (default for paid Go, Plus, and Pro plans) and GPT-Live-1 mini (default for free users). Unlike the Advanced Voice Mode it replaces, GPT-Live listens and speaks at the same time, so it can be interrupted naturally, backchannel while you talk, and perform live translation while you are still speaking. It delegates complex queries to GPT-5.5 in the background. It runs on iOS, Android, and chatgpt.com, and since September 10, 2026 it is also available to developers as gpt-live-1 in the OpenAI API.
Can GPT-Live translate in real time?
Yes, and in measured testing it never broke the interpreter role across 27 minutes of audio. Translation is invoked by prompt rather than a mode toggle: tell the assistant to "simultaneously translate whatever I say into Japanese" and it speaks a running translation while you talk. The limitation is pace, not obedience: completion lag rose from 5.7 seconds to a 10 to 14 second plateau in English to Spanish over a 452-second session.
Does GPT-Live have an API?
Yes, since September 10, 2026. GPT-Live-1 is available as the model gpt-live-1 at the v1/live/sessions endpoint, priced at $0.05 per minute of conversation for the voice layer, with the backend model that handles reasoning billed separately. This changed from the July 8 ChatGPT debut, when GPT-Live was ChatGPT-only. OpenAI also ships a separate purpose-built translation model, gpt-realtime-translate, at $0.034 per minute with 70+ input languages and 13 output languages.
How much delay does GPT-Live add when translating?
Measured through the API on September 15, 2026: a median of 11.72 seconds to finish delivering a sentence in English to Spanish, rising through the session from 5.7 seconds at minute zero to a 10 to 14 second plateau by minute three. On identical audio, OpenAI's own gpt-realtime-translate measured 2.65 seconds, Gemini 3.5 Live Translate 1.51 seconds and LiveLingo 0.94 seconds, all flat across the same 452 seconds rather than drifting.
Why does GPT-Live struggle with Japanese translation?
It falls behind far enough to stop catching up. Only 17% of Japanese sentences were translated within 30 seconds of being spoken, against 98% for the same model into Spanish, the median output landed 50.7 seconds late, and 20 of 90 sentences were never translated at all. The result reproduced across two runs. OpenAI's dedicated gpt-realtime-translate covered 98% at 3.33 seconds flat on the same audio, so this is specific to prompting a conversational model rather than a limitation of OpenAI's Japanese.
Is GPT-Live free?
In ChatGPT, yes, with a tier split. GPT-Live-1 mini is the default voice model for free accounts; paid Go, Plus, and Pro plans get the larger GPT-Live-1. No new subscription was introduced, and voice usage limits vary by plan without published minute caps. The API is separate and metered at $0.05 per minute for the voice layer plus backend model usage.
What languages does GPT-Live support for translation?
OpenAI has not published a language list. Launch materials say voice is optimized for "most spoken languages" and disclose accent and fluency gaps for others. Measured behavior differs sharply by language: the same model covered 98% of Spanish sentences within 30 seconds but only 17% of Japanese. OpenAI's developer-facing translation model, gpt-realtime-translate, documents 70+ input languages and 13 output languages: English, Spanish, Portuguese, French, German, Italian, Russian, Chinese, Japanese, Korean, Hindi, Indonesian, and Vietnamese.
How does GPT-Live compare to Gemini Live for translation?
Different architectural bets, and both are now measurable. GPT-Live is a full-duplex conversational assistant that translates on request. Google's Gemini 3.5 Live Translate is a dedicated streaming speech-to-speech model. On identical 452-second audio, Gemini held a flat 1.51 second completion lag in English to Spanish and 1.92 seconds into Japanese, while gpt-live-1 rose to a 10 to 14 second plateau in Spanish and covered only 17% of Japanese sentences within 30 seconds. A dedicated translation path holds a constant offset instead of accumulating delay.
How do you switch back to Advanced Voice Mode from GPT-Live?
Open ChatGPT voice settings and select Advanced Voice Mode. GPT-Live replaced it as the default on July 8, 2026, but OpenAI kept the old mode selectable for users who need features GPT-Live does not support, such as video and screen sharing.
How do you stop GPT-Live from interrupting or saying "mhmm"?
You cannot fully disable the backchannel behavior in the consumer app. The workaround is an explicit instruction at the start of the session, such as "only translate what I say, add nothing else." In measured API testing that instruction held completely, with zero role breaks across 27 minutes, so the interpreter role itself is reliable even though conversational interjections remain possible in ChatGPT. If the behavior blocks your use case, Advanced Voice Mode remains selectable in voice settings.
Which tool should you use for live conversation translation right now?
For short exchanges, GPT-Live is good and free inside ChatGPT. For continuous speech such as a meeting, a lecture, or a long call, use a purpose-built path. Measured on identical 452-second audio, completion lag was 0.94 seconds for LiveLingo, 1.51 for Gemini 3.5 Live Translate and 2.65 for OpenAI's own gpt-realtime-translate, all flat, versus a rising 10 to 14 seconds for prompted gpt-live-1. For translated calls to regular phone numbers, none of the OpenAI surfaces dial; that is a different product category.
9. Sources
- OpenAI. Introducing GPT-Live. OpenAI blog, July 8, 2026. openai.com
- OpenAI Developers. GPT-Live-1 model reference (API availability, endpoint, pricing). developers.openai.com
- OpenAI. GPT-Live system card. OpenAI Deployment Safety, July 2026. deploymentsafety.openai.com
- OpenAI. ChatGPT Voice (help center). help.openai.com
- TechCrunch. OpenAI releases new voice models for more natural live conversations, July 8, 2026. techcrunch.com
- MarkTechPost. OpenAI Releases GPT-Live and GPT-Live-1 mini: Full-Duplex Voice Models That Delegate Deeper Reasoning to GPT-5.5, July 8, 2026. marktechpost.com
- MacRumors. OpenAI Introduces GPT-Live to Make ChatGPT Voice Feel Like a Real Conversation, July 8, 2026. macrumors.com
- The Decoder. ChatGPT can now listen and talk at the same time, July 2026. the-decoder.com
- Simon Willison. Introducing GPT-Live (notes from preview access), July 8, 2026. simonwillison.net
- OpenAI Developers. gpt-realtime-translate model reference and realtime translation guide. developers.openai.com
- Artificial Analysis. Speech to Speech Quality Index (GPT-Live-1 ranking and time to first audio). artificialanalysis.ai
- LiveLingo Research. GPT-Live-1 Interpreter Test, September 15, 2026. Per-clip data under CC-BY 4.0. livelingo.io/research/gpt-live-1-interpreter-test
- LiveLingo Research. Real-Time Voice Translation Benchmark 2026. livelingo.io/research/benchmark-2026. Dataset DOI: 10.5281/zenodo.21250032
Updated September 16, 2026 to reflect GPT-Live-1 reaching the OpenAI API on September 10, 2026, and to add the first independent measurements of its translation behavior. Model names, tier mapping, endpoint, and pricing verified against OpenAI's developer documentation on September 16, 2026. Measurements are of the API model and not of the ChatGPT consumer experience; OpenAI may change tiers, limits, language coverage, and model behavior quickly.