Google’s developer blog, in a post dated September 15, 2026, describes its latest audio models for building real-time voice applications.
What to know
- Gemini 3.8 Live and 3.8 Live Extended Thinking are native speech-to-speech models for the Live API. Google says 3.8 Live can perform tasks while keeping the conversation going, and that Extended Thinking, aimed at complex requests, ranks #1 on Artificial Analysis’ Speech-to-Speech leaderboard.
- Google lists asynchronous function calling, visual context from live inputs, precise handling of codes such as confirmation or claim numbers, and 97+ languages.
- Google lists prices of $0.005 per minute for audio input and $0.018 per minute for audio output.
- Gemini 3.5 Transcribe is a speech-to-text model for low-latency transcription. Google reports an average word error rate of 4.0% (streaming) and 2.6% (non-streaming), 85+ languages, custom vocabulary of up to 1,000 terms and a mode that removes filler words.
- Google says the Live models are released in the Gemini API and Google AI Studio. The post does not say whether they are preview or generally available, so check the docs.
Why it matters
Voice agents are a growing prompt-writing job. For comparison, see OpenAI’s GPT-Live update and our AI voice prompts.
Source: Google blog. We summarise in our own words; check the source for the full details.