Google launches Gemini 3.8 Live and 3.8 Live Extended Thinking: thinking while speaking, 97 languages

Start with how the two models differ. Gemini 3.8 Live targets scale and cost efficiency, leaning on fluid conversational intelligence plus visual grounding; Google cites second place in the Speech Agent Arena and a price point it describes as highly competitive against other frontier models. Gemini 3.8 Live Extended Thinking targets high-complexity work, and the key phrase in the announcement is reasoning and speaking simultaneously: it acknowledges a request with an early verbal cue like Let me check that, then narrates progress out loud while running multi-step background tasks, instead of freezing the conversation until an answer is ready. The demos Google cites include turning raw sketches plus live voice feedback into working React components, and coordinating multi-step bookings with asynchronous function calls. For anyone who has used a voice assistant, the difference is intuitive: the old pattern was speak-then-wait, the new one is talk while working.

The benchmark section deserves a closer read, because voice models are measured quite differently from text models. Gemini 3.8 Live Extended Thinking scores 82.6 on Artificial Analysis' Speech to Speech Quality Index, which Google says is first overall; 68.6% on tau-Voice, which measures agentic task completion, and 35.1% on Sierra's tau-Voice-banking benchmark; and 97.7% on Big Bench Audio for reasoning. Google also says that on ServiceNow's EVA-Bench the two models push the Pareto frontier for complex workflows by balancing accuracy against conversational quality, noting the run was done on the Live API in Gemini Enterprise Agent Platform. Reading those numbers takes a little patience: 82.6 is a composite of how human the interaction feels and whether it keeps up, 68.6% and 35.1% are completion rates for whether the job actually got done, and 97.7% is raw reasoning. They describe three different things, and blending them together leads to the wrong conclusion.

Several capability details matter in engineering terms. Gemini 3.8 Live processes visual input in near real time, so the model can take in what the camera sees while you keep talking; it automatically detects and switches between 97 supported languages mid-conversation, with no manual toggle; and it executes tool calls and API requests in the background without interrupting the conversation, so it can acknowledge a request verbally while the task finishes. Google's examples include guiding employee onboarding live and playing chess with near real-time visual context. On the developer side, Google names Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents as platforms that already enable voice-driven interfaces through the Gemini Live API, handling the messy real-time media streaming underneath so developers can focus on the experience, with Salesforce, Genspark and Lumeris listed as partners. All generated audio also carries a SynthID watermark, consistent with Google's approach to provenance for generated content.

Set against this week's context, the position becomes clearer. First, cadence: 3.7 Flash three weeks ago, 3.8 Flash and 3.8 Flash Cyber on September 2, and 3.8 Live plus 3.8 Live Extended Thinking on September 15 — three releases in a month, a long way from the roughly quarterly rhythm of 2025. Second, competitive direction: coverage such as the-decoder read this launch as a direct answer to OpenAI's GPT-Live-1, played as comparable capability at a fraction of the cost, which is a play for developer volume rather than for the top of a leaderboard. Third, honest benchmark reading: 82.6 for a voice model cannot be compared with text-model scores, and 35.1% on tau-Voice-banking says that in a real banking workflow, voice agents are nowhere near hands-off. For teams building voice products, the most valuable number here may be that 35.1%, because it tells you where the deepest pit is.