Skip to content
Today
Sections

Build Reports disagree8 publishers 2 min read Published Updated

Gemini 3.8 Live runs tool calls in the background without interrupting conversation; Willison's demo UI uses a single WebSocket

Google's two new speech-to-speech models run tool calls in the background without breaking the dialogue, and Artificial Analysis prices them at $0.84 and $3.50 per hour of input audio on one benchmark subset.

The Engineer · Build desk

Send a correction

Illustration accompanying Gemini 3.8 Live runs tool calls in the background without interrupting conversation; Willison's demo UI uses a single WebSocket
Generated illustration

What happened

  • Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September 2026, two speech-to-speech models that Simon Willison described as a similar shape to OpenAI's GPT-Live family.
  • Both models detect and switch between 97 languages mid-conversation, execute tool calls and API requests in the background while they keep talking, and process visual input in near real time.
  • On Sierra's banking leaderboard for customer-service tasks, Extended Thinking resolves 35.1%, against 32.0% for GPT-Live-1 Astra and 16.5% for xAI-Realtime.
  • Artificial Analysis's cost per hour of input audio on the Big Bench Audio subset puts 3.8 Live at $0.84 and Extended Thinking at $3.50, against $4.80 for Grok Voice Think Fast 2.0 and $5.83 for GPT-Live-1 Astra.
  • Simon Willison had OpenAI's GPT-6 Astra Extra High read Google's documentation and build him a browser client for holding voice conversations with the two models.

Why it matters

  • What it limits constraint Budgeting from the $0.84 and $3.50 figures imports Big Bench Audio's ratio of model reasoning to caller speech. A support queue where the agent thinks more per minute of audio pays above that average.
  • Who has to decide decision The two tiers sit 4.2 times apart in hourly cost and 6.6 points apart in speech quality. Anyone running both has to decide which turns are worth routing to the reasoning tier.
  • What's now possible capability A client that is one WebSocket and an AudioContext lets a team instrument its own capture, playback and interrupt timings instead of waiting for an SDK release to expose them.
  • Who's exposed exposure Every second of audio these models emit carries a SynthID watermark, so recordings a business republishes or archives are detectable as model-generated.

Google's own phrasing for the new capability is real-time reasoning, visual-context handling and background task execution without interrupting the dialogue [1]. Neither launch account carries a latency number, so a team rebuilding its voice budget around this has to get the numbers from its own traces.

The transport is documented. Willison's client connects to the BidiGenerateContent WebSocket endpoint and uses a Web Audio API AudioContext for both capture and playback [12]. "The implementation uses no libraries," Willison wrote [14]. Interrupting the model mid-sentence works in it [15]. Capture, playback scheduling and barge-in therefore sit in application code.

Now the two tiers. Extended Thinking scores 82.6% on Artificial Analysis's Speech to Speech Quality Index and the standard model 76.0% [5], a gap of 6.6 points [18]. Per hour of input audio, Extended Thinking costs 4.2 times the standard tier [17]. Measured against GPT-Live-1 Astra's hourly figure, the cheap tier is 6.9 times cheaper [25]; officechai.com called it roughly a sixth of the price [22].

Against GPT-Live-1 Astra (Medium) the quality lead is 1.1 points on the speech index [19] and 0.7 points on Artificial Analysis's agentic benchmark [20], while Extended Thinking's hourly cost is 40% lower [21].

For an hourly rate to transfer, the traffic has to resemble the subset it was measured on. The denominator is hours of input audio and the numerator covers whatever the model thinks and says in between, so a queue that reasons more per minute of caller speech will not land on the benchmark average. Extended Thinking is the tier sold on multi-step reasoning [9]. dev.to reports that the announcement material it saw did not include exact pricing or a full regional rollout schedule [26].

The closest thing to a deployment number is Sierra's banking leaderboard, and it leaves 64.9% of tasks unresolved [24]. dev.to arrives at the same point in general terms: the practical value will depend on how a business designs its workflows, integrations and escalation paths [23].

Availability splits by tier. 3.8 Live goes out through the Gemini API, Google AI Studio and Search Live, with enterprise access through Gemini Enterprise, and Extended Thinking adds Gemini Live plus the Workspace apps Docs, Gmail and Keep for AI Pro and Ultra subscribers [6]. dev.to reports the Gemini Enterprise for Customer Experience route as still forthcoming, with enterprise access running through private previews [8]. Google names LiveKit, Pipecat and Agora as platform partners and Salesforce and ServiceNow among the enterprise customers building on the models [10].

What to watch

  • A published Google rate card for the Live API, which would replace Artificial Analysis's per-hour estimate with billable per-token terms.
  • Whether the Gemini Enterprise for Customer Experience route ships, and what escalation hooks it exposes to a support stack.
  • Independent latency figures: time to first audio, and tool-call turnaround while background reasoning holds the floor.

Clarity's read

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

Reality

Evidence60
Adoption22
Hype gap+30
Incentives72
Confidence64

Perspective Coverage

8 publishers
Builder
Builder 48%
Operator
Operator 29%
Investor
Investor 23%
Why these scores

Claim ledger

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

  1. [1]

    Google describes the models as capable of real-time reasoning, visual-context handling and background task execution without interrupting the dialogue.

    Reported SupportedView cited source
  2. [2]

    Both Gemini 3.8 Live models can detect and switch between 97 languages mid-conversation, execute tool calls and API requests in the background while continuing to talk, and process visual input in near real time.

    Reported SupportedView cited source
  3. [3]

    Google released Gemini 3.8 Live and 3.8 Live Extended Thinking on 15 September 2026, two speech-to-speech models that are a similar shape to OpenAI's GPT-Live family.

    Reported SupportedView cited source
  4. [4]

    On Artificial Analysis's cost-per-hour-of-input-audio measure using the Big Bench Audio subset, Gemini 3.8 Live costs $0.84 an hour, 3.8 Live Extended Thinking $3.50, Grok Voice Think Fast 2.0 $4.80 and GPT-Live-1 Astra $5.83.

    Reported SupportedView cited source
  5. [5]

    On Artificial Analysis's Speech to Speech Quality Index, Gemini 3.8 Live Extended Thinking scores 82.6%, GPT-Live-1 Astra (Medium) 81.5%, Grok Voice Think Fast 2.0 (High) 81.3% and the standard Gemini 3.8 Live 76.0%.

    Reported SupportedView cited source
  6. [6]

    3.8 Live rolls out through the Gemini API, Google AI Studio and Search Live with enterprise access via Gemini Enterprise; Extended Thinking gets the same developer and enterprise rollout plus a consumer release inside Gemini Live and Google Workspace Docs, Gmail and Keep for Google AI Pro and Ultra subscribers.

    Reported SupportedView cited source
  7. [7]

    On Artificial Analysis's tau-Voice benchmark, 3.8 Live Extended Thinking scores 68.6%, GPT-Live-1 Astra 67.9% and Grok Voice Think Fast 2.0 56.5%.

    Reported SupportedView cited source
  8. [8]

    dev.to reports that Gemini Enterprise customers get private previews and that the models coming to Gemini Enterprise for Customer Experience is explicitly described as forthcoming.

    Reported SupportedView cited source
  9. [9]

    Google positions the standard model as a scalable, cost-efficient option for live dialogue and the Extended Thinking version for more complex work that benefits from increased intelligence and multi-step reasoning.

    Reported SupportedView cited source
  10. [10]

    Google says it is partnering with platforms including LiveKit, Pipecat and Agora, along with enterprise customers including Salesforce and ServiceNow, to build voice agents on the new models.

    Reported SupportedView cited source
  11. [11]

    Simon Willison pointed GPT-6 Astra Extra High at Google's documentation and had it build a web UI for trying the new models, with model and voice preset selection, an optional system prompt, and a voice conversation in the browser.

    Reported SupportedView cited source
  12. [12]

    The UI connects to the wss://generativelanguage.googleapis.com BidiGenerateContent WebSocket endpoint and uses a Web Audio API AudioContext for both capture and playback.

    Reported SupportedView cited source
  13. [13]

    All audio output from the models carries Google's SynthID watermark, its technology for helping detect AI-generated content.

    Reported SupportedView cited source
  14. [14]

    "The implementation uses no libraries," Willison wrote.

    Reported SupportedView cited source
  15. [15]

    The browser UI includes the ability to interrupt the model while it is talking.

    Reported SupportedView cited source
  16. [16]

    On Sierra's tau-cubed Banking Leaderboard, which tests how well a voice agent resolves customer-service style banking tasks, 3.8 Live Extended Thinking leads with 35.1%, against 32.0% for GPT-Live-1 Astra and 16.5% for xAI-Realtime.

    Reported SupportedView cited source
  17. [17]

    Extended Thinking costs about 4.2 times the standard tier per hour of input audio.

    Derived Supported
  18. [18]

    The Speech to Speech Quality Index gap between Extended Thinking and the standard model is 6.6 points.

    Derived Supported
  19. [19]

    Extended Thinking's lead over GPT-Live-1 Astra (Medium) on the Speech to Speech Quality Index is 1.1 points.

    Derived Supported
  20. [20]

    Extended Thinking's lead over GPT-Live-1 Astra on the tau-Voice benchmark is 0.7 points.

    Derived Supported
  21. [21]

    Extended Thinking's hourly input-audio cost is 40% lower than GPT-Live-1 Astra's.

    Derived Supported
  22. [22]

    officechai.com wrote that the standard Gemini 3.8 Live model is roughly a sixth of the price of GPT-Live-1 Astra for a good chunk of the same conversational ability.

    Reported SupportedView cited source
  23. [23]

    dev.to writes that the practical value will depend on how a business designs its workflows, integrations and escalation paths.

    Reported SupportedView cited source
  24. [24]

    Extended Thinking's 35.1% score on Sierra's banking leaderboard leaves 64.9% of tasks unresolved.

    Derived Supported
  25. [25]

    Gemini 3.8 Live's hourly cost is 6.9 times lower than GPT-Live-1 Astra's, closer to a seventh than a sixth.

    Derived Supported
  26. [26]

    dev.to reports that Google has not provided exact pricing or a full regional rollout schedule in the material it saw.

    Reported ContestedView cited source

Sources

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

  1. blog.google

    3 articles · September 15, 2026

    Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
  2. dev.to

    2 articles · September 18, 2026

    Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe

  3. officechai.com

    1 article · September 17, 2026

    Google Releases Gemini 3.8 Live-Extended Conversational Model, Claims Better Performance Than Rivals At Lower Price
  4. runtimewire.com

    1 article · September 15, 2026

    Google ships Gemini 3.8 Live to keep voice agents talking while tools work
  5. simonwillison.net

    2 articles · September 15, 2026

    Gemini Live audio
  6. testingcatalog.com

    2 articles · September 15, 2026

    Google rolled out Gemini 3.8 Live and Extended Thinking
  7. the-decoder.com

    4 articles · September 15, 2026

    Apple brings a fully revamped Siri built on Google's Gemini, but not to the EU
  8. thenewstack.io

    2 articles · September 15, 2026

    OpenAI’s voice model doesn’t think. That’s the point.

Topics and entities

Topics

Entities