GoHighLevel Voice AI Is Moving to Gemini 3.8 Live: What to Test Before the Switch

Short answer: Google is shutting down the gemini-3.1-flash-live-preview model on November 17, 2026 and names Gemini 3.8 Live as its replacement. HighLevel speech-to-speech Voice AI agents can already run on Gemini 3.8 Live, and release roundups report that agents still on the older model move over automatically on November 2. Your prompts stay the same, but the model answering your callers changes, so spend 15 minutes test-calling each agent before and after the switch.

What is changing, and what is confirmed

• Confirmed by Google: the Gemini API retires gemini-3.1-flash-live-preview and an older native audio preview model on November 17, 2026. Google names gemini-3.8-live as the replacement, and the old endpoints stop working rather than redirecting.
• Reported for HighLevel: speech-to-speech Voice AI agents can use Gemini 3.8 Live, and agents on the older preview model migrate automatically on November 2. This comes from release roundups, not a HighLevel changelog entry we could find, so check the model setting on your agent and any notice in your account.
• Not affected: agents that do not use a speech-to-speech Gemini model.

For an automatic migration you do not need to rebuild anything. The reason to care is that a different model can sound and behave differently on the same prompt.

What may behave differently on calls

Google's notes for developers list real differences between the two models. HighLevel handles the technical settings for you, but the effects can still show up on a call:

• Tool calls run differently by default. Gemini 3.8 Live runs function calls asynchronously by default. On a voice agent, that is the moment it checks a calendar or books an appointment, so test booking end to end.
• Proactive audio is always on. The agent may respond to pauses and background sound differently, so test interruptions and silence.
• Affective dialogue was removed. Tone matching may feel different, so listen to how it handles an upset caller.

A 15-minute test plan for each agent

1. Note which model the agent uses now and save a copy of its prompt and settings.
2. Switch the agent to Gemini 3.8 Live ahead of the deadline if your account lets you, so the change happens on your schedule, not mid-week.
3. Make five test calls: a normal booking, a reschedule, a pricing question, an interruption mid-sentence, and a caller who goes quiet.
4. Check the results in the account: appointment created on the right calendar, contact fields filled, call summary and transfer working.
5. Read the transcripts for anything the agent now says that it should not, and tighten the prompt if needed.
6. Repeat two of the calls after November 2 to confirm nothing moved.

When to get help

If your voice agent books appointments or takes payments, a bad week costs more than the test. We set up and manage voice agents as part of AI Front Desk, and we keep existing builds tested through changes like this on the GHL Care Plan. For what makes a voice agent sound natural in the first place, see AI answering that sounds human, and for everything else HighLevel announced this month, the LevelUp 2026 announcements page.

Yaakov (Koby) Oranski, Founder and CEO of Yotomations

Certified GoHighLevel Admin and Partner. Koby designs CRM migrations, two-way integrations, billing systems and AI agents for service businesses, and writes from the engagements described on this site.

Published 2026-10-27