01
Overview
Google launched two closely related real-time voice models on September 15, 2026: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. They share the same Live API ecosystem, but Google positions them for different workloads.
The short version: choose Gemini 3.8 Live when fast, fluid conversation and cost-efficient scale matter most. Evaluate Extended Thinking when the agent must solve complex, multi-step tasks while keeping the spoken interaction alive.
02
The core difference
Gemini 3.8 Live is the general-purpose model. Google describes it as built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding.
Gemini 3.8 Live Extended Thinking adds deeper, configurable reasoning for complex tasks. It can reason in the background while speaking, acknowledge a request before the work is finished and narrate progress through a multi-step process.
That makes Extended Thinking less like a simple “better quality” tier and more like a different reasoning mode for workflows where the agent needs time to think without making the user wait in silence.
04
Where Extended Thinking differs
Extended Thinking supports configurable thinking for complex multi-step reasoning. Google demonstrates it coordinating multi-step bookings, transforming rough sketches into React components from live voice feedback and continuing to speak while background tasks run.
For a workflow that requires several dependent decisions, ambiguous constraints or long-running tools, that behavior can be valuable. For a straightforward FAQ or transactional agent, it may be unnecessary complexity.
05
Pricing
Google’s September 15 developer announcement lists the new Live models at $0.005 per minute for audio input and $0.018 per minute for audio output.
Google’s launch materials do not justify assuming every possible workload has identical total cost. Extended Thinking can perform more reasoning, and production systems may incur other infrastructure or modality costs. Verify the current pricing documentation and measure cost per completed task rather than comparing only the headline audio rate.
06
Benchmarks
Google says Extended Thinking ranks #1 overall on Artificial Analysis’ Speech-to-Speech Quality Index with a score of 82.6. It also reports 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking benchmark and 97.7% on Big Bench Audio.
Google says standard Gemini 3.8 Live placed second on Speech Agent Arena.
These are useful signals, not a universal winner declaration. The results are presented in Google’s launch material and may not represent your exact language, latency, tool stack or business workflow.
07
Choose Gemini 3.8 Live when...
Use standard Live as the first model to test when the task is conversational but relatively direct: customer FAQs, simple appointment flows, guided onboarding, voice search, basic sales qualification or applications where response speed and scale are more important than extended reasoning.
It is also the sensible baseline for benchmarking because it lets you determine whether deeper reasoning actually improves task completion enough to justify using it.
08
Choose Extended Thinking when...
Test Extended Thinking when the agent needs to coordinate several tools, resolve complex constraints, reason over visual and spoken context, plan multi-step actions or explain progress while a difficult task is running.
Examples include complex booking, technical troubleshooting, financial or operational workflows with several validation steps, and voice-driven software-building or analysis tasks.
09
Do not choose from benchmark rank alone
Voice-agent performance is multidimensional. A high benchmark score does not guarantee lower latency, fewer tool failures or a better user experience in your application.
Measure interruption handling, time to first useful response, total task time, tool-call success, correction rate, abandonment and cost per successful outcome.
10
Rollout differences matter
The API availability of the two models is similar at launch, but Google’s consumer and enterprise rollout differs.
Standard 3.8 Live is rolling out in Search Live and private-preview enterprise surfaces. Extended Thinking is rolling out in Gemini Live and selected Workspace experiences, while enterprise access also includes private-preview and coming-soon surfaces.
Do not infer that access in one Google product means the same model is available in every account or region.
11
A practical selection process
Start with standard Gemini 3.8 Live on a representative set of real tasks. Record task completion, latency and user corrections. Then run the same set against Extended Thinking for the subset of tasks that require more reasoning.
If Extended Thinking materially improves completion or reduces failed tool loops, use it for those workflows. If not, standard Live may provide the simpler production choice.
A hybrid router may eventually make sense: direct easy tasks to standard Live and reserve Extended Thinking for complex jobs. That is an architectural inference, not a Google-prescribed requirement.
12
FAQ
Is Extended Thinking a separate model?
Google presents Gemini 3.8 Live Extended Thinking as a distinct Live variant with deeper configurable reasoning.
Is Extended Thinking always better?
No. It is designed for more complex tasks. A simpler workload may benefit more from standard Live’s scale- and efficiency-oriented positioning.
Do both models support tool calls?
Google documents asynchronous function calling across the new Live model family.
Do both support visual context?
Yes, Google lists visual grounding/live visual context among the family’s core capabilities.
Which one is cheaper?
Google’s launch developer post quotes the same audio input/output rates for the new Live models. Total task cost can still differ with workload behavior and should be measured in production-like tests.
Which should I start with?
For most straightforward voice-agent workloads, standard Live is a reasonable baseline. Test Extended Thinking specifically where deeper multi-step reasoning is needed.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.