Google launches Gemini 3.5 Live Translate for 70+ languages

• Google shipped Gemini 3.5 Live Translate as a low-latency speech-to-speech model for 70+ languages, according to Google's launch thread and the official announcement.
• The rollout spans three surfaces at once: public preview in the Gemini Live API and AI Studio, global availability in Google Translate on Android and iOS, and a Google Meet private preview for select Workspace customers this month, per Google's rollout summary.
• Google is pitching continuous translation rather than turn-taking. Google's demo thread says the model translates as you speak without pauses, while Google AI Studio's feature list adds multilingual auto-detection, pitch and pacing preservation, and noise filtering.
• The developer-facing model is gemini-3.5-live-translate-preview, with audio input, translated audio output, optional transcripts, and no tool support, according to the model page and the Live Translation guide.
• Google's pricing page lists the paid tier at $3.50 per 1M input tokens and $21 per 1M output tokens, which it also translates to about $0.0368 per audio minute.

Google published the launch post, the model card, and a full Live Translation guide on day one. The docs are unusually concrete: they spell out the exact model slug, the PCM audio formats, and the translationConfig switch that turns a Live API session into an interpreter instead of a general-purpose voice agent.

Google split the launch across consumer, developer, and enterprise entry points.

• Google Translate app: live translate is rolling out globally on Android and iOS across 70+ languages, per Google's app rollout thread.
• Gemini Live API and AI Studio: developers get public preview access, per Google's rollout summary and the model page.
• Google Meet: speech translation enters private preview this month for select business Google Workspace customers, with broader rollout later this year, according to Google's Meet preview note.

Google's own thread also makes the product split explicit: the same model is being positioned as an end-user Translate feature, a programmable API surface, and a Meet capability in waiting.

Google's main technical pitch is that the model does not wait for a speaker to finish a turn.

• Continuous translation: Google's demo thread says it generates translated speech as you speak, without the pause pattern of turn-based systems.
• Language auto-detection: Google AI Studio's feature list says it can detect multilingual inputs inside a single session.
• Voice preservation: the official announcement says translated speech keeps intonation, pacing, and pitch.
• Noise handling: Google AI Studio's feature list calls out robust noise filtering for loud environments.

The launch post adds one more useful detail: Google says the model balances waiting for context against translating immediately, and usually stays only a few seconds behind the speaker.

The consumer app shipped with one feature Google did not foreground in the API materials: Android gets a new listening mode that plays translated audio through the phone earpiece instead of requiring headphones, according to Google's app and Meet update.

That same post also nails down the near-term product path:

• Now: Google Translate on Android and iOS.
• This month: Meet private preview for select Workspace business customers.
• Later this year: broader Meet rollout, per Google's Meet preview note.

For engineers, that means Google is treating the API launch and the first-party product rollout as the same model story, not separate research and product tracks.

The Live Translation guide draws a hard product boundary between Live Translation and the regular Live Agent flow.

• Model slug: gemini-3.5-live-translate-preview, per the model page.
• I/O shape: audio in, translated audio out, plus optional input and output transcripts, per the model page and guide.
• Session config: translation is enabled with translationConfig, including targetLanguageCode and echoTargetLanguage, per the guide.
• Audio transport: input is raw 16-bit PCM at 16 kHz mono, output is raw 16-bit PCM at 24 kHz mono, and Google recommends 100 ms chunks, per the guide.
• Token limits: 131,072 input tokens and 65,536 output tokens, per the model page.
• Capabilities it does not have: the docs say Live Translation is audio-only, translation-only, and does not support tools or instructions, unlike Live Agent sessions, per the guide.
• Pricing: the pricing page lists a free tier, then paid pricing of $3.50 per 1M input tokens and $21 per 1M output tokens, or about $0.0368 per minute of audio.

That last part is the real engineering tell. Google did not just ship a demoable voice feature. It shipped a narrowly-scoped real-time translation pipeline with its own model slug, config path, transport contract, and price sheet.