Google's September 15 API changelog records general availability of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The documented distinction concerns real-time audio interaction and background reasoning needs. Businesses evaluating voice agents should test both conversational behavior and operational consequences.
Evaluate the audio experience
Use realistic background noise, interruptions, different speaking speeds, and clarification requests. Measure whether the agent preserves the caller's intent after an interruption rather than simply producing natural-sounding speech.
Do not select a model solely from a latency claim. The relevant measure is how long it takes to resolve the caller's task.
Evaluate tool use independently
If the agent looks up an account or changes an appointment, test identity verification, permissions, and confirmation of the exact action. Spoken fluency must not become a substitute for authorization.
Provide a clear handoff when the agent cannot resolve a request. Test whether a human receives useful context without unnecessary sensitive information.
Control recording and disclosure
Review consent, disclosure, data retention, and the handling of transcripts under applicable rules. Synthetic voice capabilities do not remove the need to tell users what kind of service they are interacting with where required.
Pilot with a bounded task such as answering approved service questions before enabling consequential actions. A successful voice agent is understandable, interruptible, and recoverable. Its most important feature may be knowing when to stop and transfer the call.
Sources & further reading
Reviewed October 26, 2026. Product availability and regulatory guidance can change. Vendor performance claims are not independent Rovio AI benchmarks.
Found this useful? Share it.