Talking and thinking: Google's background bet with Gemini 3.8 Live

On 15 September 2026 Google unveiled Gemini 3.8 Live and its Extended Thinking variant, two native speech-to-speech dialogue models designed to handle real-time conversation. The most significant technical change is not how smooth the voice sounds, but the ability to call external tools and run processing in the background while the user is still speaking. As TechRepublic observed in its analysis, this split means the end of a spoken answer no longer necessarily coincides with the completion of the underlying operation. The models, which support dynamic switching across 97 languages and near-real-time visual input, also embed SynthID watermarking to identify generated audio.

On performance metrics, the Extended Thinking version took first place in Artificial Analysis's Speech to Speech Quality Index with a score of 82.6. It is a private benchmark with its own methodology and volatile rankings, and the gap to the rest of the field is thin: according to figures reported by Insider Monkey, OpenAI's GPT-Live-1 Astra sits at 81.5 points and xAI's Grok Voice Think Fast 2.0 at 81.3. Insider Monkey itself noted that a margin of just 1.1 points may reflect how crowded the race has become rather than any structural technology gap. On internal benchmarks, Google claims 97.7 per cent on Big Bench Audio and 68.6 per cent on τ-Voice, which nonetheless drops to 35.1 per cent on Sierra's harder τ-Voice-banking scenario. With no independent third-party check, those last numbers have to be read as one-sided vendor claims.

The initial rollout covers both developer APIs and integration into commercial services such as Search Live and the Gemini app, but several grey areas remain around operating costs and infrastructure stability. The trade press reports rates of 3 dollars per million audio input tokens and 12 dollars per million output tokens (roughly 0.005 and 0.018 dollars a minute) for the standard version, yet Google's official pricing page, checked on 19 September 2026, still does not list these specific entries, nor does it put a number on the extra cost of Extended Thinking's reasoning tokens. On top of that, the entire Live API remains in preview with no defined date for general availability, and no measured data has been published on actual latency or on error rates across languages.

Choosing not to publish measured latency or accuracy across the 97 supported languages leaves the main question open: whether the separation between the flow of conversation and the background processing can hold up in everyday use, outside controlled benchmarks. It will be the stability of the real infrastructure, rather than tenths of a point in private tests, that decides whether invisible voice interaction is a concrete step forward or just an acceleration to avoid falling behind.