Gemini 3.8 Live: What Google’s New Voice Agent Models Actually Do

What actually shipped

On September 15, 2026, Google released two new models into the Gemini Live API and Google AI Studio: Gemini 3.8 Live and a second, reasoning-focused variant called Gemini 3.8 Live Extended Thinking. Both are live now, API-only — there is no self-hosted or open-weight version of either. They're built for the specific job of a spoken, real-time agent: something that listens, can see, calls tools mid-conversation, and talks back without the round-trip of a text pipeline bolted onto a voice layer.

The capabilities that are actually new

The headline feature is asynchronous function calling: the model can execute a tool call or an API request in the background while the conversation keeps moving, instead of going silent while it waits on a response. Paired with that is real-time visual input processing (the model can be shown something mid-call, not just told about it), stated alphanumeric precision for reading back things like order numbers or confirmation codes, and incremental content updates that stream both audio and structured data together rather than one blocking the other. Extended Thinking is the same model family with a configurable reasoning budget for the cases that need a multi-step plan before the model opens its mouth.

None of these are unique concepts — background tool use and streaming responses are the direction the whole voice-agent field has been moving — but this is Google's specific claim about what its own models now do, via the original announcement.

The benchmarks, and where they stop being flattering

Gemini 3.8 Live Extended Thinking ranks first on Artificial Analysis' Speech-to-Speech Quality Index at 82.6, and scores 97.7% on Big Bench Audio, a reasoning-over-audio eval. Those are strong, general-purpose numbers. On the same announcement, Google also reports 68.6% task completion on τ-Voice, a general agentic-voice benchmark from Sierra — and 35.1% on Sierra's narrower τ-Voice-banking variant, which tests the model on actual banking-support tasks rather than open-ended agentic ones.

That's close to a 2-in-3 drop from the general eval to the domain-specific one, and it's Google's own reported number, not a third-party knock. It's the detail most likely to get flattened into "Gemini 3.8 Live tops the leaderboard" coverage. A model that is genuinely first-place on general speech quality can still fail two-thirds of the time on a regulated, high-stakes vertical task — which is exactly the kind of task a lot of "production-grade voice agent" pitches are actually for. The gap is the reason to test against your own task before reading the top-line score as a green light.

The one benchmark Google didn't win outright: Gemini 3.8 Live placed second in the Speech Agent Arena, a human-preference evaluation. The announcement doesn't name what took first, so treat that one as "strong, not the leader" rather than filling in a guess.

Pricing is per-minute, not per-token

Gemini 3.8 Live bills audio the way a phone system would, not the way an LLM API usually does: $0.005 per minute of audio input and $0.018 per minute of audio output, which Google states is derived from a $3-per-million-input-token / $12-per-million-output-token rate underneath. If you're used to estimating voice-agent cost in text tokens, convert to minutes-of-call-time first — it's the unit that will actually show up on your bill, and it's also the unit that makes a support-line cost projection comparable across vendors.

Who you'd actually build with, at launch

For the wider set of speech and audio tools already in the directory, see text-to-audio tools and every Gemini model ranked. If you're weighing voice generation (TTS/narration) against a conversational voice agent like this one, those are different jobs — see what AI voice generation actually is for the distinction.