Google Unveils Gemini 3.8 Live and Extended Thinking Language Models

Google has released two new language models, Gemini 3.8 Live and Live Extended Thinking, aimed at improving speech-to-speech capabilities in AI-driven workflows.

The models are designed to power complex tasks with intuitive conversations without awkward silences, a feature that has proven difficult to achieve so far.

Gemini 3.8 Live achieved an index score of 82.6 on Artificial Analysis' speech test, outperforming its competitors in the Speech Agent Arena.

The model also demonstrated impressive results in other tests, including a 97.7 percent accuracy rate on Big Bench Audio and a 68.6 percent score on τ-Voice.

Extended Thinking showed even more promise, with scores of 35.1 percent on Sierra's τ-Voice banking benchmark and 97.7 percent on Big Bench Audio.