Google releases Gemini 3.8 Live voice AI models, undercutting OpenAI on price

Google releases Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, native speech-to-speech models designed for real-time voice agents. The Extended Thinking model claims the top spot on the Artificial Analysis Speech to Speech Quality Index with a score of 82.6, surpassing OpenAI's GPT-Live-1. Google undercuts OpenAI on price, charging $0.005/min for input and $0.018/min for output. The models support background tool execution, visual context, and 97 languages, and are available via the Gemini API and Google AI Studio.

Gemini 3.8 Live release

• ▪The Gemini 3.8 Live models extend Google's Gemini Audio family, which was expanded in August 2026 with the release of Gemini 3.5 Transcribe.
• ▪Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, as its most advanced native speech-to-speech models for real-time voice agents.

Speech-to-speech benchmark rankings

• ▪Gemini 3.8 Live Extended Thinking scored 68.6% on the τ-Voice agentic task completion benchmark, 35.1% on Sierra's τ-Voice-banking benchmark, and 97.7% on Big Bench Audio.
• ▪The standard Gemini 3.8 Live model placed fifth on the Artificial Analysis Speech to Speech Quality Index with a score of 76.0, while Gemini 3.1 Flash Live High scored 71.5.
• ▪Gemini 3.8 Live Extended Thinking achieved first place on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6, surpassing GPT-Live-1 and Grok Voice.
• ▪In blind Speech Agent Arena conversations, human testers preferred the older Gemini 3.1 Flash Live model, which leads with an Elo of 1096, over Gemini 3.8 Live at 1083 and Extended Thinking at 990.

API pricing undercuts OpenAI

• ▪Google's pricing of $0.005 per minute for audio input and $0.018 per minute for output undercuts OpenAI's GPT-Live-1, which costs $0.05 per minute.
• ▪An hour of voice conversation costs approximately $1.38 with Google's standard Gemini 3.8 Live model, compared to at least $3.00 with OpenAI's GPT-Live-1.
• ▪Google charges $0.005 per minute for audio input and $0.018 per minute for audio output for both Gemini 3.8 Live models via the Live API.

Real-time reasoning capabilities

• ▪The Gemini 3.8 Live models support near real-time visual context processing, allowing voice agents to understand visual inputs alongside spoken user prompts.
• ▪The Gemini 3.8 Live models feature automatic language detection and mid-conversation switching across 97 supported languages while maintaining accent consistency.
• ▪Gemini 3.8 Live Extended Thinking supports real-time, multi-step reasoning in the background while simultaneously speaking, using early verbal cues like 'Let me check that' to acknowledge prompts.

Background tool execution

• ▪All audio generated by the Gemini 3.8 Live models carries Google DeepMind's imperceptible SynthID watermark to assist in detecting misinformation.
• ▪The Gemini 3.8 Live models execute asynchronous third-party software tool and API calls in the background, allowing audio responses to stream continuously without conversational interruption.

Production API availability

• ▪Google has partnered with developer platforms including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents to support real-time media streaming infrastructure for the Gemini 3.8 Live models.
• ▪The Gemini 3.8 Live models are available for production use in the Gemini Live API and Google AI Studio, with no self-hosted open-weights option.

Debatable claims

• ▪Google should offer an open-weights version of its Gemini Live models
• ▪AI developers should be required to watermark all generated audio