Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two speech-to-speech models built along the same lines as OpenAI’s GPT-Live family. The number everyone will quote is 82.6 on Artificial Analysis’ Speech to Speech Quality Index, first place for Extended Thinking. Fine. That leaderboard will have a new occupant by December, and no serious product decision should hang on a couple of index points.
The feature worth your attention is duller and much more useful: these models execute tool calls and API requests in the background while the conversation keeps flowing.
The dead air problem
Anyone who has shipped a voice agent knows the failure mode. The user asks something that needs a lookup. The model goes quiet. Two seconds, three, four, while a function call resolves against some sluggish internal API, and by then the user has started talking again, or said “hello?”, or hung up. The workarounds are all cosmetic. Canned “let me check on that for you” audio. Fake typing sounds. Filler tokens generated to burn clock.
Moving the tool call under the speech stream removes the need for any of it. That’s the difference between a demo that impresses a room and something a company will put in front of paying customers who are already annoyed about their broken order.
Same logic applies to live visual input. Pointing a phone camera at a router and talking through it only works if the model isn’t stalling every time it needs to consult a manual.
The language claim deserves side-eye
Ninety-seven languages, switchable mid-conversation, is a real engineering achievement and also a marketing number. Language counts always are. The top dozen will be excellent because that’s where the training data lives. Somewhere down the list, quality falls off a cliff, and no vendor publishes where. If you’re deploying into Tagalog or Amharic, test it yourself before you trust the count.
The mid-conversation switching is the genuinely interesting bit. Code-switching is normal human behavior in most of the world, and handling it without a language-detection round trip is the kind of thing that makes an agent feel less like a phone tree.
A small detail that says a lot about 2026
Simon Willison built a working test client for the new models by pointing GPT-6 Astra Extra High at Google’s documentation. A rival lab’s model, given nothing but docs for an API released that morning, produced a functional web UI with model selection, voice presets, and system prompt support.
Two things follow from that. Google’s launch documentation is good enough to be machine-readable in the most literal sense. And the realtime voice API surface has converged hard enough across vendors that a model trained before this release already knew the shape of it.
What Google hasn’t published is what Extended Thinking’s extra reasoning costs in latency during an actual call. Ranking first on a quality index means little if the thinking adds 800ms to every turn. That’s the number I want before either model goes anywhere near a customer.