Google is introducing a new pair of language models: Gemini 3.8 Live and Live Extended Thinking. When it comes to speech-to-speech, these models take the lead. Beyond that, Google must figure out how to regain its competitive edge against general-purpose LLMs.
The models are now available via the Gemini API and Google AI Studio. The goal of Gemini 3.8 Live and Live Extended Thinking is to power AI-driven reasoning and agentic workflows without awkward silences. The combination of complex tasks and intuitive conversation has proven difficult so far, but with an index score of 82.6 on Artificial Analysis’ speech test, 3.8 Live is leading the pack.
Furthermore, the results for Extended Thinking are particularly impressive. It achieved a score of 68.6 percent on τ-Voice, 35.1 percent on Sierra’s τ-Voice banking benchmark, and 97.7 percent on Big Bench Audio. The less powerful Gemini 3.8 Live came in a close second in the Speech Agent Arena.
Speaking and thinking at the same time
Google follows the concept of OpenAI’s GPT-Live, which also combines longer workloads with natural-sounding language. Here, too, phrases like “Let me just check this” are important for the credibility of the AI conversation. OpenAI places a strong emphasis on this, incorporating plenty of “hm”s and “ah”s to break the silence.
Gemini 3.8 Live can also view images practically in real time. It can automatically switch between 97 supported languages and execute tools and API calls in the background.
Earlier this month, Google released Gemini 3.8 Flash, featuring a context window of 1 million tokens, a standard feature these days.
Ecosystem and availability
Platforms such as Agora, LangChain, LiveKit, Pipecat, and Vercel support the Live API. Salesforce, Genspark, and Lumeris are still testing the models. Anyone who wants to know whether a real person is speaking can, in theory, determine this by looking for a SynthID watermark.
Gemini 3.8 Live is also available in private preview within Gemini Enterprise. Extended Thinking is coming to Gemini Live, Docs, Gmail, and Keep for Google AI subscribers.
Uncertain frontier
With its mix of efficient Flash models now in place, Google must also have plans to hit back at the frontier labs. It seems that no Pro model is immediately forthcoming. For some reason, Google’s models are currently entirely occupied with speed and affordability – we hesitate to see any reason other than struggling development internally for this near half-year of non-Pro releases. It has meant the likes of SpaceXAI and Meta have pushed past the capabilities of Gemini family models.