Gemini 3.8 Live for Voice Agents: Latency, Price per Minute, and When to Leave Twilio

Google shipped gemini-3.8-live and gemini-3.8-live-extended-thinking on 15 September 2026. Native speech-to-speech, Live API, listed around $0.005 per minute input and $0.018 per minute output, with background tool calls while audio is in flight. Distinct from Gemini 3.8 Flash (chat/coding) and from 3.5 Transcribe.

If you have been assembling voice agents from Whisper or Deepgram, a chat model, and ElevenLabs or Cartesia, this is the first generally pitched SKU in months that makes "delete the three-hop pipeline" a serious option. It is not automatic, and it is not a reason to rip out telephony.

For the older Omni Flash path, see building a sub-second voice agent. For turn-taking physics, see voice agent latencies and WebRTC.

What changed in the architecture

Classic hop:

1. Stream audio to STT (200 to 500ms)
2. Wait for LLM first token (400 to 1200ms)
3. Stream TTS (200 to 400ms)

Best case you are still over a second of dead air. Users say "hello?" and the agent double-talks.

Live SKUs take raw audio in and raw audio out. Tool calls can run in the background so the model does not have to stop speaking to fetch an order. Extended Thinking is the variant that is allowed to pause longer when the lookup is actually hard. Use it on collections and claims, not on "what are your hours."

Price per minute, not per million tokens

Voice invoices in minutes. At $0.005 in and $0.018 out, a five-minute support call with roughly equal talk time is on the order of $0.06 for model audio, before telephony, recording, and tools.

Compare that to a three-hop stack:

• STT: often $0.006 to $0.015 / min
• LLM: highly variable; a chatty agent can blow past $0.05 / min in tokens
• TTS: $0.015 to $0.04 / min depending on the voice

Live can win on the model line. It does not win if you still pay Twilio for the leg, still record for QA, and still run a separate LLM for after-call summary. Add those up before you tell finance the voice bot is now "two cents a call."

When to leave the three-hop stack

Move a route to 3.8 Live when:

• Time-to-first-audio is your actual complaint, not NLU accuracy
• The conversation is English-first (97-language switching is marketed; verify the languages you care about on a real call)
• Tools are few, well typed, and safe to run while the caller is listening
• You can accept Google as the voice brain for that line of business

Keep STT plus LLM plus TTS when:

• You need a specific branded voice that Live cannot match
• Compliance requires a transcript you already trust from a named STT vendor
• The agent must run on a model Google does not host (Opus 5 on the language, Live only on the mouth: possible, but you are back to hops)
• PCI or healthcare recording rules make a new audio data-flow a legal project, not a sprint

Telephony stays. SIP trunks, DIDs, queueing, warm transfer, and "press 1" are not Gemini features. Live is the brain and the mouth. Twilio, LiveKit, or your CCaaS is still the wire. See voice agent latency on telephony.

Design rules we would ship with

1. Barge-in first. If the caller talks over the agent and the agent keeps going, you will get one-star calls. Test this on day one, not after the demo.
2. Tools in the background, results in the next sentence. "I am pulling that up" is allowed. Silence is not.
3. Extended Thinking only on a named intent. Default Live for greetings and FAQs. Thinking for exceptions. Otherwise you pay latency on every turn.
4. Transcript anyway. Even if you do not use STT in the loop, persist a transcript for QA and evals. You cannot improve what you cannot read.
5. Shadow, then cut over one queue. IVR FAQ first. Password reset second. Never the collections line first.

Frequently Asked Questions

Is 3.8 Live the same as 3.8 Flash? No. Flash is a text/agent coding SKU. Live is speech-to-speech. Do not put Flash on a phone call and do not put Live on a code-review agent.

Can we keep our ElevenLabs voice? Then you are not on native Live output. You can still use Live for understanding and a separate TTS for the mouth, but you have reintroduced a hop. Measure it; do not assume it is "almost as fast."

What about EU and India data residency? Confirm where audio is processed and stored on the Live API before you point a European or Indian number at it. Voice is personal data. The SKU launch does not answer that for you.

How do we eval a voice agent? Text evals on transcripts plus a latency budget (p50 and p95 time-to-first-audio, barge-in success rate, talk-over rate). If you only score the transcript, you will ship a polite agent that feels drunk on the phone.

---

FoundrySoft builds production voice agents with real telephony, not just a browser mic demo. See AI voice agents or get in touch.