Google’s real-time voice line runs from the 2018 Duplex phone agent to Gemini 3.8 Live, generally available since September 15, 2026. The new models keep talking while reasoning and tool calls run in the background. Google’s documents publish no measured response time for them and list three different counts of the languages they speak.
1. The September 2026 launch
On September 15, 2026, Google made two models generally available in the Live API, its real-time interface for voice and vision: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking (Gemini API changelog). The launch post, by Alisa Fortin and Thor Schaeff of Google DeepMind, puts asynchronous function calling at the top of the feature list. Tool and API calls run in the background while the model keeps streaming audio to the user.
Extended Thinking adds background reasoning. Google’s guide to thinking in the Live API says the model plans and calls tools in the background and speaks short updates while it works. Its example update is “Checking flight options now.” Because the model can speak several times before a task is finished, the guide tells developers to watch a status field instead of the end of a spoken turn.
Google has used spoken filler this way before. In May 2018 its research blog described Duplex, a system that phoned businesses to book appointments in a few narrow domains. Duplex put “hmm” and “uh” sounds into its pauses, and the team wrote that these told the other person the system was still processing. Gemini’s updates do the same job with a sentence about the task instead of a syllable.
OpenAI sells its own version of the same split. Its model page describes GPT-Live 1 as a full-duplex voice model that hands reasoning and tool use to a backend agent, with the backend configured and billed separately.
Google exposes the listening voice and the background work as one model ID inside one Live API session, and its documents describe what the client sees without naming the component that reasons. We think this packaging is the most consequential choice in the launch. One model is easier to adopt than OpenAI’s two parts, but it hides the piece a team would most want to inspect or swap.
2. Google’s audio research before Gemini
Duplex was a pipeline. Google’s 2018 post says a recurrent neural network, built with TensorFlow Extended, read the output of Google’s speech recognizer along with audio features, the conversation so far and the task parameters. Speech came from a concatenative text-to-speech engine combined with a synthesis engine using Tacotron and WaveNet. Where a quick reply was expected, it could fall back on faster, lower-confidence models, and in those cases it answered in under 100 ms. Tasks it could not finish went to a human operator, and the post kept Duplex to closed domains: it could not hold a general conversation.
In May 2019, Ye Jia and Ron Weiss introduced Translatotron, a single sequence-to-sequence model that turned speech in one language into speech in another without an intermediate text step. They called it the first such end-to-end model “to the best of our knowledge” and noted that its translations still trailed a conventional cascade.
In August 2021, Neil Zeghidour and Marco Tagliasacchi described SoundStream, a neural audio codec that runs in real time on a smartphone CPU. At 3 kbps it outperformed Opus at 12 kbps. A codec like this turns audio into discrete tokens, and a language model can learn to predict them.
AudioLM (paper first posted on arXiv on September 7, 2022) did that. It chained Transformer models over semantic tokens from w2v-BERT and acoustic tokens from SoundStream, and given a few seconds of speech it generated a continuation. Human raters told real speech from AudioLM continuations 51.2% of the time, against the 50% a coin toss would give, though a classifier the team trained caught the generated speech with 98.6% accuracy.
On April 20, 2023, Google merged the Brain team from Google Research with DeepMind to form Google DeepMind, the unit that now builds Gemini. When Kyutai launched in Paris in November 2023, its launch announcement said Zeghidour, AudioLM’s last author, had joined from Google DeepMind, where he started and led a team working on generative audio. Kyutai later released Moshi, an open full-duplex model (see the Kyutai profile).
Google has not published how much of this research sits inside today’s Live models. The Gemini 3.8 Audio model card says only that the models are based on Gemini 3 Pro and points to that model’s card for architecture.
- May 2018Duplex books appointments by phone, with “hmm” and “uh” in its pauses
- May 2019Translatotron translates speech to speech without intermediate text
- August 2021SoundStream codec: 3 kbps beats Opus at 12 kbps
- October 2022AudioLM continues speech from SoundStream and w2v-BERT tokens
- April 2023Brain and DeepMind merge into Google DeepMind
- August 2024Gemini Live in the Gemini app, Android, English
- December 2024Multimodal Live API arrives with Gemini 2.0 Flash
- May 2025Native audio models and asynchronous function calls in the Live API
- December 2025Gemini 2.5 Flash Native Audio, also rolling out in Gemini Live and Search Live
- March 2026Gemini 3.1 Flash Live preview
- September 2026Gemini 3.8 Live and Extended Thinking generally available; Live Avatar in Gemini Enterprise
3. Gemini Live in the app and the Live API
Gemini Live, the voice mode in the Gemini app, came first. Google introduced it on August 13, 2024 for Gemini Advanced subscribers on Android, in English. It could be interrupted mid-answer and kept talking with the phone locked, and it shipped with ten new voices. The post promised iOS and more languages “in the coming weeks.” The Gemini app release notes list the iPhone app with Gemini Live in November 2024. In May 2026 a redesign put Live inside the chat, so users can switch between speaking and typing; it rolled out worldwide on Android and iOS.
Developers got the Live API on December 11, 2024, when Google announced a Multimodal Live API with Gemini 2.0 Flash that took real-time audio and video-streaming input and could combine several tools. The changelog records a billed public preview model on April 9, 2025, with session resumption, context compression and configurable voice activity detection. On May 20, 2025, Google launched asynchronous function calls and released Live API models with native audio output. A June 2025 post by Ankur Bapna and Tara Sainath described what those models could do. Proactive audio ignored background speech, affective dialog adapted to the user’s tone, the model could reason during the conversation, and it covered more than 24 languages.
The app and the API began sharing models in December 2025. The updated Gemini 2.5 Flash Native Audio reached the Gemini API on December 12 (changelog), and Google’s post about it made it generally available on Vertex AI, kept it in preview in the Gemini API, and began rolling it out in Gemini Live and Search Live. Google reported 71.5% on ComplexFuncBench Audio and instruction adherence of 90%, up from 84%. The Gemini 3.1 Flash Live preview followed on March 26, 2026, and again went to developers, enterprises and consumers at the same time. Google said it brought Search Live to more than 200 countries and territories and reported 90.8% on ComplexFuncBench Audio and 36.1% on Scale AI’s Audio MultiChallenge with thinking on. Google says all of its audio carries a SynthID watermark.
The September 15 launch post covers the Gemini API and Google AI Studio. It does not say whether Gemini 3.8 Live now runs Gemini Live in the app, and it does not mention Vertex AI. Other voice models arrived in the same weeks: Gemini 3.5 Transcribe in August, Gemini 3.8 Flash TTS and Flash-Lite TTS on September 22 (changelog), and on September 24 a Live Avatar for Gemini 3.8 Live in Gemini Enterprise, with custom avatars limited to allowlisted enterprise customers.
On January 22, 2026, TechCrunch, citing Wired, reported that Google DeepMind was bringing on Alan Cowen, Hume AI’s CEO, and about seven engineers to work on Gemini’s voice features, under a licensing agreement. Andrew Ettinger, who took over as Hume’s CEO, said Google has a “non-exclusive right to certain technologies” (see the Hume AI profile).
| 2.5 Flash Native Audio | 3.1 Flash Live | 3.8 Live | 3.8 Live Extended Thinking | |
|---|---|---|---|---|
| Model ID | gemini-2.5-flash-native-audio-preview-12-2025 | gemini-3.1-flash-live-preview | gemini-3.8-live | gemini-3.8-live-extended-thinking |
| Released | December 12, 2025 | March 2026 | September 15, 2026 | September 15, 2026 |
| Non-blocking tools | Supported | Not supported | Default | Required |
| Proactive audio | Configurable | Not supported | Always on | Always on |
| Affective dialog | Supported | Not supported | Removed | Removed |
| Status, November 10, 2026 | Shutdown from November 17, 2026 | Shutdown from November 17, 2026 | Stable, no shutdown date | Stable, no shutdown date |
Gemini 3.8 Live also takes away settings developers used to have. Proactive audio, which lets the model stay silent when speech is not meant for it, was a setting on the 2.5 model and is permanently on in 3.8; switching it off returns an error. Affective dialog, which adapted replies to the user’s emotional tone, is supported only on the 2.5 model and has been removed from the 3.8 models (capabilities page). Non-blocking tool calls, which the March 2026 preview did not support at all, are the default in 3.8 Live. The removals surprised us more than the additions: affective dialog was one of the features the June 2025 post described, and the model pages note its removal without giving a reason.
4. How the Live API handles background work
In a delegation design, one model holds the conversation, listening and deciding when to speak, while slower work such as reasoning, database lookups and API calls happens behind it and comes back into the conversation when it is ready. OpenAI builds this from two named parts, a voice model and a backend agent. Google offers it at three depths, from a flag on a function to a dedicated model.
Non-blocking function calls
Since May 2025, a developer can mark a function NON_BLOCKING. According to the tools guide, the function then runs while the user keeps talking to the model. When the client sends the result back, it chooses how the model should use it: interrupt what it is doing and report right away, wait until it has finished what it is doing, or say nothing and use the result later. A function that forwards the request to a separate agent or text model is a delegation the developer builds. The client runs the tools itself, since the Live API does not handle tool responses automatically.
Gemini 3.8 Live
The 3.8 Live model page makes non-blocking calls the default and keeps blocking calls for backward compatibility. (Google’s general tools guide still describes blocking as the default and asks for each function to be flagged, and the thinking guide’s comparison table says the standard model waits for tool execution before speaking.) The page lists “interleaved reasoning” among the model’s capabilities but does not accept thinking settings, and calls 3.8 Live the default choice for voice agents “without reasoning-induced delays.”
Gemini 3.8 Live Extended Thinking
The Extended Thinking model page adds background reasoning with a depth of low, medium or high. Every function call must be non-blocking; a blocking declaration returns an error. The thinking guide asks developers to flag every function non-blocking, set the reasoning depth, and stop treating the end of a spoken turn as the end of the task. The server can keep reasoning or calling tools after it marks a turn complete, so the client watches a field called interaction_status, which reads IN_PROGRESS during background work and IDLE when the whole task is finished. Google says the standard model has a fixed latency profile, while Extended Thinking lets the developer set the depth and uses spoken updates to cover slow tools.
The Extended Thinking page says the model “introduces background reasoning during live audio sessions” but not whether that reasoning runs in the same network that speaks, in a second model, or somewhere else. The general capabilities page says that when a user interrupts, the server discards the ongoing generation and any pending function calls, and reports the calls’ IDs to the client. The thinking guide does not say whether background reasoning survives an interruption.
The two companies also bill the split differently. OpenAI charges for the voice session and, separately, for whatever backend model and tools the developer configures (GPT-Live 1 model page). Google bills one model by the token, and its pricing table labels the output price as including thinking tokens (Gemini API pricing). A Google developer tunes how much one model thinks, while an OpenAI developer picks the backend model that does the thinking.
The trade-off we would weigh is control against convenience. Non-blocking calls let a team send work to any backend it chooses, at the cost of building that routing itself. Extended Thinking does the routing, but every tool has to be non-blocking, and the reasoning runs somewhere Google’s documents do not describe.
5. Languages, prices and limits
Languages
Google’s documents disagree on how many languages the Live API speaks. The launch post says Gemini 3.8 Live supports 97+, the Live API overview says users can converse in 70, and the capabilities page says 99 without a separate list per model. The Live Avatar post says 97. Per the capabilities page, native audio models choose the language themselves and can switch mid-conversation; developers cannot pass a language code and can only restrict languages through system instructions. For comparison, Google described the June 2025 native audio preview as supporting over 24 languages. A language count is one of the first figures a buyer checks, so we would want Google to say which of its three numbers it stands behind.
The neighbouring models have counts of their own. Google says Gemini 3.5 Live Translate covers more than 70 languages and Gemini 3.5 Transcribe 85 or more, with a reported average word error rate of 4.0% when streaming and 2.6% when not (launch post).
Prices
On the Gemini API pricing page, Gemini 3.8 Live and Extended Thinking share one price: $3.00 per million audio input tokens, which Google converts to about $0.005 a minute, and $12.00 per million audio output tokens, about $0.018 a minute. Image and video input costs $1.00 per million tokens, or about $0.002 a minute. Both models have a free tier. The audio rates are the same as for the December 2025 native audio preview; the text rates went up, from $0.50 to $0.75 per million input tokens and from $2.00 to $4.50 per million output tokens.
| Model | Audio in | Audio out | Text in, out (per 1M tokens) | Billed by |
|---|---|---|---|---|
| Gemini 3.8 Live and 3.8 Live Extended Thinking | $3.00 per 1M tokens, about $0.005 a minute | $12.00 per 1M tokens, about $0.018 a minute | $0.75, $4.50 | Tokens; the output price includes thinking tokens |
| Gemini 2.5 Flash Native Audio (December 2025 preview) | $3.00 per 1M tokens | $12.00 per 1M tokens | $0.50, $2.00 | Tokens |
| OpenAI GPT-Live 1, for reference | $0.05 per session minute, billed per second | Backend model billed separately | Session time, plus backend usage | |
Limits and availability
The capabilities page caps audio-only sessions at 15 minutes and audio-with-video sessions at 2 minutes, and says session management can extend both without a fixed limit. Native audio models get a 128k-token context window. The model card gives 128K tokens of input, 64K tokens of output and a knowledge cutoff of January 2025. Audio goes in as 16-bit PCM at 16 kHz and comes out at 24 kHz; images go in at up to one frame per second (overview).
The API runs over a stateful WebSocket. For apps that connect straight from the client, Google recommends short-lived tokens over standard API keys, and WebRTC is available through partners. The launch post names Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel and Vision Agents as integrations. The older models are being retired: the deprecations table gives November 17, 2026 as the earliest shutdown date for both the December 2025 native audio preview and the 3.1 Flash Live preview, with gemini-3.8-live as the replacement.
6. Outside benchmarks
Google’s September 15 launch post said Extended Thinking was ranking first on Artificial Analysis’ speech-to-speech leaderboard. That is Google’s statement, and the current standings are on Artificial Analysis’ leaderboard. The changelog describes the base model as engineered for sub-second conversational latency, but Google publishes no measured figure or method.
Two independent papers have measured earlier Gemini Live models. Full-Duplex-Bench v3, posted in April 2026 by researchers at National Taiwan University and NVIDIA, tests tool use when the user’s speech contains real-world disfluencies. Between Gemini Live 2.5 and 3.1, the time from the end of the user’s speech to the factual answer fell from 7.26 s to 4.25 s, and the share of tasks with every tool call exactly right rose from 0.49 to 0.54. GPT-Realtime, tested alongside, scored 0.60 at 6.89 s.
Gemini Live 3.1 also responded at a natural moment in 78% of scenarios, against 92% for 2.5. The paper says it produced no speech in the other 22%, although in 86% of those silent cases it still called the tools. It opened 31.7% of its responses with a filler sentence that said nothing, against 8.9%. The paper treats a lower filler rate as better: fillers reduce perceived wait time, but in its tests they often overlapped with the user’s speech. Extended Thinking speaks updates on purpose, and Google’s example names the task in progress, which is more than a content-free filler does. In our view this is the main open question about 3.8 Live: whether its updates carry information or just fill time. The paper predates 3.8 Live, so it cannot answer it.
InteractionBench, posted on November 2, 2026 by AveraLabs, ran gemini-3.1-flash-live-preview through the Live API between August 17 and 31, 2026, alongside six other systems. Gemini Live scored 77% on timing, 36 of 50 on spoken tasks, and passed 46 of 52 groundedness probes (88%): turns whose correct reply depends on what was said earlier in a conversation of 22 to 24 turns. It was also the only system that whispered when asked to on most requests, 8 of 10, judged from the audio signal.
The authors list limits on these numbers. Hosted systems are timed where the client hears them, and because Gemini Live reports no server timestamps, they cannot tell how much of its timing reflects network delay. The groundedness judge belongs to the same model family as Gemini Live, and the same model generated the test conversations. The user’s speech is synthesized.
The 2018 Duplex post reported that its fillers made calls feel more natural in user studies. Neither paper here tests Gemini 3.8 Live, and neither asks how listeners take Extended Thinking’s spoken updates.
Google got to this design in steps: non-blocking tool calls in May 2025, the default in 3.8 Live, then background reasoning in Extended Thinking, all sold as one model on one bill. What a team would want to check before putting it on a phone line is what Google has not published: a measured response time, how often Extended Thinking speaks an update, and what happens to background work when the caller interrupts.
Sources
- Google, Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe, September 15, 2026
- Gemini API docs: Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, Thinking in the Live API, Live API tools, Live API capabilities, Live API overview, pricing, deprecations, changelog, read November 10, 2026
- Google DeepMind, Gemini 3.8 Audio model card, September 2026
- Google: Gemini Live (August 13, 2024), Gemini 2.0 (December 11, 2024), Gemini 2.5 native audio (June 3, 2025), improved Gemini audio models (December 2025), Gemini 3.1 Flash Live (March 26, 2026), Gemini 3.8 Live with Live Avatar (September 24, 2026), Gemini app release notes
- Google Research: Duplex (May 8, 2018), Translatotron (May 15, 2019), SoundStream (August 12, 2021), AudioLM (October 6, 2022) and its paper
- Google, Google DeepMind announcement, April 20, 2023; CMA CGM, Kyutai launch, November 17, 2023; TechCrunch, Hume AI team report, January 22, 2026
- OpenAI, GPT-Live 1 model page; Kyutai, Moshi paper; Artificial Analysis, speech-to-speech leaderboard
- Lin et al., Full-Duplex-Bench v3, April 2026; He et al., InteractionBench, November 2026
Next: NVIDIA: open full-duplex speech from a GPU company, and the licences that come with it