Skip to content
Today
Sections

Product Reports disagree3 publishers 2 min read Published Updated

Gemini 3.8 Live keeps talking while it runs the tool call in the background

Google's two new voice models process speech and reasoning at the same time and can fire tool calls without pausing the conversation. Published scores put the stronger one top of a speech-to-speech index and lowest on banking tasks.

The Product Desk

Send a correction

Illustration accompanying Gemini 3.8 Live keeps talking while it runs the tool call in the background
Generated illustration

What happened

  • Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, pitching both at the latency problem that has kept voice agents scripted and shallow.
  • Both models execute third-party tool and API calls in the background, so the agent keeps conversing at a normal pace while it works on the task it was just given.
  • On Google's own numbers, Extended Thinking scored 82.6 on the Artificial Analysis Speech to Speech Quality Index, ahead of GPT-Live-1-Astra and Grok Voice Think Fast 2.0.

Why it matters

  • What it costs cost Output audio bills at 3.6 times input, so a script that makes the agent talk more costs more than one that makes it listen more, and the difference lands on whoever owns the call minutes.
  • What it limits constraint Because reasoning tokens are billed on top for Extended Thinking, a per-minute rate cannot be used to forecast what the higher-scoring model will cost; that number has to come from measured calls.
  • Who has to decide decision Teams picking between the cheap predictable model and the top scorer now have to make that call per use case, because the same model's scores move by tens of points across domains.
  • What it sets up precedent Verbal filler shipped as a documented feature sets the expectation that agents cover their own thinking time with speech, which QA has to start scoring alongside accuracy.

"Let me check that" is now a listed product feature. Google says the models use early verbal cues like it to acknowledge a prompt in a more lifelike way [4]. Contact centres have been running that line for decades under a plainer name. SiliconANGLE's account of Google's post describes near-real-time reasoning without a latency measurement, so how much of the wait actually shrank is not in the published material [12].

The scores are Google's own, and SiliconANGLE published them with the aside "if Google is to be believed" [8]. Inside a single model the range is wide: 97.7% on Big Bench Audio, 68.6% on T-Voice, and 35.1% on T-Voice-banking [7]. Banking sits 33.5 points below general T-Voice and 62.6 points below Big Bench Audio [16]. On that set, 64.9% of attempts were not successes [17]. The standard model placed second on Speech Agent Arena and first on ServiceNow's EVA-Bench [6].

A ten-minute call split evenly between the two sides works out at 5 x $0.005 plus 5 x $0.018, or about 11.5 cents of audio [19]. Extended Thinking bills reasoning tokens on top of that, along with inputs such as video and documents [10].

For the person who has to deploy this, the two models arrive in different places. Both are live in the Gemini API and Google AI Studio, with an enterprise private preview in Gemini Enterprise and Search Live [20]. Extended Thinking also reaches Workspace through Docs, Gmail and Keep for subscribers, plus the Gemini Live apps [21]. Existing voice stacks connect through partners including LiveKit, Pipecat, Agora, Vercel, Fishjam and Vision Agents [13]. The models handle 97 languages, detect the language automatically and can switch mid-conversation [11], and generated audio carries an invisible SynthID watermark that Google said can be used to detect misinformation [14].

Google AI's launch post said the models "let you speak, collaborate, and execute tasks seamlessly, meaning conversing with AI just got a lot more natural" [15]. Two things sort whether that holds for a given use case. The first is whether the tool call can finish while the caller keeps talking, or whether its answer gates the agent's next sentence; background execution only pays in the first case, and in the second the caller hears the filler and then waits anyway. The second is what a wrong answer at that step costs the person on the line, because a mis-stated account balance delivered fluently is a different failure from a mis-stated opening hour. For regulated, transactional work of the kind T-Voice-banking is meant to stand in for, completion rate on last week's real calls is a better guide than any index score.

What to watch

  • Whether Google or an independent tester publishes end-to-end latency figures for either model.
  • Whether the Gemini Enterprise and Search Live private preview opens up with a published per-minute rate for Extended Thinking.
  • Whether an independent run of T-Voice-banking reproduces the score Google reported.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence45
Adoption
Not scored
Hype gap+30
Incentives70
Confidence55

Perspective Coverage

3 publishers
Builder
Builder 40%
Operator
Operator 40%
Investor
Investor 20%
Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

  1. [1]

    Google LLC launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, aiming at the latency problem associated with voice-based AI agents.

    Reported Supported Source: SiliconANGLEView cited source
  2. [2]

    Google billed the two models as its most advanced voice processing models released so far, capable of near-real-time reasoning and simultaneous speech-and-thought processing.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  3. [3]

    Google said the models are designed to execute third-party software tool and API calls in the background while maintaining a natural pace of conversation, so the agent can keep talking while working on a task it was just assigned.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  4. [4]

    Google said the models can use early verbal cues such as "let me check that" to acknowledge a user's prompt in a more natural, lifelike way.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  5. [5]

    Gemini 3.8 Live Extended Thinking achieved a score of 82.6 on the Artificial Analysis Speech to Speech Quality Index, surpassing GPT-Live-1-Astra and Grok Voice Think Fast 2.0.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  6. [6]

    Gemini 3.8 Live ranked second on the Speech Agent Arena benchmark and first on ServiceNow Inc.'s EVA-Bench.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  7. [7]

    Gemini 3.8 Live Extended Thinking scored 68.6% on T-Voice, 35.1% on T-Voice-banking and 97.7% on Big Bench Audio.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  8. [8]

    SiliconANGLE reported the benchmark results as Google's, writing that the models "boast top-tier benchmark results, if Google is to be believed".

    Reported Supported Source: SiliconANGLEView cited source
  9. [9]

    Google said the standard Gemini 3.8 Live model costs $0.005 per minute for audio inputs and $0.018 per minute for outputs.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  10. [10]

    Google said the Extended Thinking model also charges for reasoning tokens and for additional inputs such as video and documents.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  11. [11]

    Both models support automatic language detection and mid-conversation language switching, understand and generate speech in 97 languages, and support near real-time visual grounding.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  12. [12]

    SiliconANGLE's report of the launch describes the models as capable of near-real-time reasoning but does not include a latency measurement.

    Reported Supported Source: SiliconANGLEView cited source
  13. [13]

    Google said developers can integrate the models through partner platforms including Vercel, Agora, LiveKit, Pipecat, Fishjam and Vision Agents.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  14. [14]

    Google said generated audio files carry an invisible SynthID watermark, which can be used to detect misinformation.

    Reported Supported Source: Google, via SiliconANGLEView cited source
  15. [15]

    A Google AI post on X said: "Gemini 3.8 Live and 3.8 Live Extended Thinking let you speak, collaborate, and execute tasks seamlessly, meaning conversing with AI just got a lot more natural."

    Reported Supported Source: Google AI (@GoogleAI), September 15, 2026View cited source
  16. [16]

    The T-Voice-banking score is 33.5 points below the general T-Voice score and 62.6 points below the Big Bench Audio score.

    Derived Supported
  17. [17]

    On T-Voice-banking, 64.9% of attempts were not scored as successes.

    Derived Supported
  18. [18]

    The published output rate is 3.6 times the input rate for the standard model.

    Derived Supported
  19. [19]

    A ten-minute call with five minutes of user audio and five minutes of model audio costs $0.115 in audio at the standard model's published rates.

    Derived Supported
  20. [20]

    Gemini 3.8 Live is available via the Gemini API and Google AI Studio, and as an enterprise private preview in Gemini Enterprise and Search Live.

    Reported Contested Source: Google, via SiliconANGLEView cited source
  21. [21]

    Gemini 3.8 Live Extended Thinking is available through the same channels and also in Google Workspace via Docs, Gmail and Keep for subscribers, and in the Gemini Live applications.

    Reported Contested Source: Google, via SiliconANGLEView cited source

Sources

3 independent publishers whose own reporting we read for this story.

  1. 9to5google.com

    2 articles · September 15, 2026

    Gemini 3.8 Live Extended Thinking powers Gemini Live, Gmail, & Keep
  2. siliconangle.com

    2 articles · September 15, 2026

    Google’s new speech model Gemini 3.8 Live supports real-time reasoning
  3. thurrott.com

    1 article · September 15, 2026

    Google Announces Gemini 3.8 Live and 3.8 Live Extended Thinking - Thurrott.com

Topics and entities

Topics

Entities