On September 15, 2026, Google announced its real-time voice interaction models "Gemini 3.8 Live" and "Gemini 3.8 Live Extended Thinking." The former emphasizes low latency and cost efficiency, while the latter focuses on background reasoning for solving multi-step problems. Both models accept audio, images, video, and text as input, and return audio and text. Developers can now access both through the Gemini API and Google AI Studio.
The difference here isn't simply about extending thinking time. Extended Thinking continues reasoning or external tool processing for the same request while verbally communicating intermediate progress. This means applications must separately track "the model has finished speaking" from "the entire job is complete." While silent waiting periods can be reduced, this creates new burdens around state management, asynchronous tools, and the cost of long sessions.
An API That Separates Speech Completion From Task Completion
Gemini 3.8 Live is primarily intended for use cases requiring quick responses, such as inquiry routing, language practice, voice search, and short device operations. It includes fixed-type reasoning woven into the conversation, but users cannot select the amount of reasoning. Function calling defaults to asynchronous, though synchronous calls can be specified for compatibility.
Extended Thinking is designed for tasks like technical support that examines multiple logs, bookings that search flights and accommodations in parallel, or educational support that verifies mathematical formulas and code. Users select a reasoning level from low, medium, or high, and external functions only accept NON_BLOCKING asynchronous execution. Specifying synchronous calls results in an error.
In Extended Thinking, turnComplete: true merely indicates the end of an utterance—overall request completion is determined by interaction_status: IDLE.
| Implementation Item | Gemini 3.8 Live | Extended Thinking |
|---|---|---|
| Reasoning | Fixed-type, woven into conversation | Background reasoning at low/medium/high levels |
turnComplete |
Normally indicates turn end and waiting | One utterance boundary; processing may continue |
| Overall request status | Managed via traditional turn boundaries | Managed via interaction_status (IN_PROGRESS/IDLE) |
| Tools | Supports both synchronous and asynchronous | Asynchronous only |
| Response while waiting | Waits until tool completion | Can provide verbal progress updates |
For example, in a travel search scenario, even if turnComplete arrives the moment the model says "I'm checking flights," the search function or comparison processing may continue. If the client returns the microphone to standby mode or finalizes the results screen at this point, it risks missing subsequent tool calls or the final answer. The connection and UI state must be maintained until interaction_status becomes IDLE.
The standard version also supports asynchronous function calling. So the distinction isn't that "the standard version cannot process in parallel." Rather, in Extended Thinking, variable reasoning and multiple utterances span a single request, changing the meaning of completion itself. The internal structure of what Google calls "parallel reasoning" and its actual degree of parallelism have not been disclosed—all that can be verified externally is this API-level behavior.
An 82.6 Score, Yet Only 35.1% for Banking Tasks
According to Artificial Analysis evaluations at the time of measurement, Extended Thinking (High) scored 82.6 on the Speech to Speech Quality Index. This breaks down into 98% for instruction following, 91.9% on Full Duplex Bench (which measures conversational simultaneity), 68.6% on τ-Voice (which measures agent task completion), and 97.7% on Big Bench Audio (which measures reasoning involving audio). Time to first audio was 1.35 seconds, with an estimated price of $3.50 per hour.
The standard version's overall score was 76.0, with τ-Voice at 30.1% and Big Bench Audio at 89.1%. However, Full Duplex Bench came in at 96.1%, time to first audio was 1.18 seconds, and the estimated price was $0.84 per hour—giving the standard version an advantage in conversational responsiveness and cost. This is not a result where Extended Thinking is superior across all voice interaction scenarios.
Google separately reports that Extended Thinking scored 35.1% on Sierra's τ-Voice-banking benchmark, which focuses specifically on banking operations. This differs from the 68.6% across all industries and cannot be directly compared as equivalent accuracy. Rather, this figure should be read as: even with natural intermediate responses and high reasoning scores, the model failed to complete roughly two-thirds of tasks in challenging banking operations.
The 97.7% Big Bench Audio score also ranked 4th at the same measurement point—a different evaluation axis from Google's claimed "overall #1." The standard version's #2 ranking on Speech Agent Arena reflects user preference, not task completion rate. Rather than lining up rankings, adoption testing needs to separately assess vocal naturalness, initial response, reasoning, and final task completion rate.
ServiceNow's EVA framework doesn't measure speech recognition or synthesis component by component, but instead evaluates how failures accumulate across an entire multi-turn conversation. While Google claims both models push the Pareto Frontier of accuracy and conversation quality, the execution environment shown in the figure is the Live API on the Gemini Enterprise Agent Platform. The same results may not hold for the general Gemini API or consumer-facing Gemini.
Longer Conversations Cost More Even at the Same Unit Price
The Gemini API pricing table groups 3.8 Live, Extended Thinking, and the previous 3.1 Flash Live Preview under the same listing. For the paid tier, per-million-token rates are: $0.75 for text input, $3 for audio input, $1 for image/video input, $4.50 for text output, and $12 for audio output. Converted to Google's time-based rates, this comes to roughly $0.005 per minute for audio input and roughly $0.018 per minute for audio output.
From this, one cannot conclude that "Extended Thinking costs the same for the same task." The Live API reprocesses and re-bills tokens from past context remaining in the current context window on each turn. This means the cost of a single response increases as the conversation grows longer. Without configuring context compression to drop older history, costs can accumulate in long customer service or diagnostic sessions in ways not visible from hourly rate conversions alone.
When transcribing audio input and output, text output charges for the generated transcription apply on top of audio fees. In both models, proactive audio is always enabled, meaning input tokens are billed while the model is listening. Since average token counts by reasoning level or actual cost per task have not been published, the practical approach is to run identical call scenarios through both the standard version and Extended Thinking and compare cost per completed task.
The free tier offers input and output at no cost, but content may be used to improve Google products. The paid tier is the category that excludes use for improvement purposes. When handling customer calls, it's worth checking data processing terms first, not just pricing.
Deployment Availability Varies by Destination
Both models share input/output limits of 131,072 tokens input and 65,536 tokens output according to API documentation, though model cards round these to 128K and 64K respectively. Both are built on Gemini 3 Pro, with a knowledge cutoff of January 2025. Google itself lists hallucination and occasional delays or timeouts as known limitations. The ability to converse in real time is no guarantee of external information accuracy or task completion. Generated audio includes a SynthID watermark.
Availability status differs by product. Developers can access both models via the Gemini API and Google AI Studio. Enterprise-facing Gemini Enterprise is in private preview, while Gemini Enterprise for Customer Experience is coming soon. Extended Thinking's rollout to Google Workspace business customers is also coming soon.
On the consumer side, 3.8 Live is being deployed to Search Live, while Extended Thinking is rolling out to Gemini Live. Within Workspace, this covers Docs for Google AI Pro/Ultra subscribers, and Gmail and Keep for all Google AI subscribers. Since the announcement did not provide a country- or language-specific timeline, the model's capability for automatic switching across 97 languages should be considered separately from when each product becomes available in Japan.
Adoption decisions should be based not on model name alone, but on conversation speed, task complexity, whether the client can accommodate state management, task completion rates by use case, and the cost of long sessions. For instantly returning short requests, the standard version is the candidate; for scenarios requiring multiple tools that take several seconds without making users wait, Extended Thinking becomes the candidate. What's needed next is the public disclosure of latency and actual costs by reasoning level, along with completion rates by language—including Japanese—based on real-world business scenarios.