Gemini 3.8 Live and its Extended Thinking variant, released by Google on September 15, listen, see, search and act without breaking the conversation. This Gemini 3.8 Live test puts both models through latency, conversational handling and real cost, head to head with GPT-Live-1 and Grok.
Key Takeaways
- Extended Thinking tops Artificial Analysis’ speech quality index at 82.6, ahead of GPT-Live-1 at 81.5.
- The base model answers in 1.18 seconds and costs $0.84 per hour of input audio, the cheapest rate on the board.
- On conversational flow, GPT-Live-1 keeps the edge: 94.9% against 91.9% for Extended Thinking.
Have an AI Sum Up This Article
ChatGPTWhat Gemini 3.8 Live changes in day-to-day use
Google now splits voice into two jobs. Gemini 3.8 Live is built for scale and cost: fluid conversation, visual grounding and real-time language detection across 97 languages. Gemini 3.8 Live Extended Thinking targets complex requests, with multi-step reasoning that does not stall the flow of the exchange.
In use, the most noticeable change is how tools run. Function calls fire in the background, so the model kicks off a search or an action and keeps talking while it runs. It fills the gap with acknowledgments along the lines of “let me check that,” then narrates progress when a task has several steps.
That is where this Gemini 3.8 Live test sees the clearest break with the previous generation. Dead air while a backend call completes is a familiar weak spot of voice bots, and moving the tool call off the critical path of the conversation addresses it at the model level instead of in application code.
The model also processes visual input almost in real time and switches languages mid-conversation. For multilingual customer service, that is the capability that counts most: a caller who moves from French to English no longer needs a new session.
The voice itself already splits opinion. On Product Hunt, user André J called the model “really good” but a bit over expressive, to the point of feeling that the AI was trying to sell him something. For a brand, that upbeat register can be tuned in the instructions, but it is worth knowing before putting the model on the phone.
On the consumer side, Extended Thinking powers Gemini Live, voice features in Gmail and Keep for Google’s AI subscribers, and Docs for AI Pro and Ultra subscribers in Workspace. The base model runs the Search Live experience in AI Mode, and Gemini Live is rolling out on the new model now.
Those Workspace features show what the model is meant to do day to day. Gmail Live answers questions about your inbox out loud while showing a transcript with clickable links to the source emails, Docs Live turns spoken ideas into a structured first draft, and Keep Live sorts a rambling voice memo into lists and notes within seconds.
Every piece of generated audio carries an imperceptible SynthID watermark meant to help detect AI-generated content. We covered the launch itself in our piece on how Gemini 3.8 Live undercuts OpenAI on voice pricing.
Our Gemini 3.8 Live test: latency, conversation and real cost
First criterion, responsiveness. The base model delivers its first audio in 1.18 seconds, against 2.99 seconds for Gemini 3.1 Flash Live, the previous generation. Extended Thinking, which reasons before it answers, stays at 1.35 seconds, level with GPT-Live-1 in its Astra version at 1.34 seconds.
Second criterion, conversational handling: pauses, turn-taking, interruptions and small verbal cues. The base model reaches 96.1% on this measure from Artificial Analysis, which runs every voice model through the same set of trials, far above the 74.3% of Gemini 3.1 Flash Live. Extended Thinking drops to 91.9%, and GPT-Live-1 Astra keeps the lead at 94.9%.
Third criterion, getting a task done. On τ-Voice, which simulates customer service exchanges, Extended Thinking completes 68.6% of tasks against 30.1% for the base model, which even trails Gemini 3.1 Flash Live at 37.7%. This Gemini 3.8 Live test therefore points to a clean split of roles between the two models.
Google adds two data points of its own. It cites 97.7% on Big Bench Audio, an audio reasoning benchmark, and reports that the base model took second place in the Speech Agent Arena, where users compare voice agents and vote for the one they prefer. For a Gemini 3.8 Live test focused on real conversations, that preference vote matters as much as any lab score.
Last criterion, real cost. An hour of input audio runs $0.84 on the base model, the lowest rate on the board, and $3.50 on Extended Thinking. On the other side, GPT-Live-1 costs $4.47 in its Sol version and $5.83 in its Astra version, with Grok Voice Think Fast 2.0 at $4.80.
Scaled to a call center, the gap becomes tangible: a hundred hours of listening cost $84 on the base model, $350 on Extended Thinking and $583 on GPT-Live-1 Astra. For teams buying at volume, that is the number in this Gemini 3.8 Live test that will weigh the most.
The official price list explains the gap: $0.005 per minute of audio input and $0.018 per minute of output, unchanged from the previous generation, plus a free tier to get started. On the free tier, content can be used to improve Google’s products, which makes it a prototyping tool rather than a place for real customer calls.
Speed is the one area where Google does not lead. Grok remains the fastest of the group, with first audio in 0.70 seconds, which matters for use cases where every fraction of a second of silence feels awkward, such as outbound sales calls.
Google documents the limits itself: possible hallucinations, a knowledge cutoff of January 2025, jailbreak resistance still being improved, and occasional slowness or timeouts. The model does not support caching, code execution, file search, structured outputs or URL context, so developers who rely on those features will need to keep them in another part of the stack. On safety, the model card states that neither variant reached a Tracked or Critical Capability Level under Google’s Frontier Safety Framework.
More articles on Horizon
- OpenAI Models Tried to Hide Their Mistakes
- DeepMind Institute Tackles the Rise of Human-Level AI
- Review a Contract With Claude Without Leaking Data
Gemini 3.8 Live against GPT-Live-1: our verdict
On measured quality, Extended Thinking edges past GPT-Live-1 on both voice quality and completed tasks, at an hourly cost roughly 40% lower. OpenAI keeps the advantage on natural flow, thanks to a model that listens and speaks at the same time.
This duel picks up where we left off when GPT-Live let ChatGPT listen and talk at the same time. Google is answering with price, a very responsive base model and distribution across its own products, while OpenAI leans on its full-duplex architecture.
Our verdict after this Gemini 3.8 Live test: adopt it for high-volume voice agents, and keep Extended Thinking for conversations that trigger real actions. The base model, fast and fluid, fits front-desk calls, information intake and simple questions.
Teams building interruption-heavy experiences, such as sales calls or coaching, should run both vendors side by side on their own recordings before committing. That is where the conversational edge of GPT-Live-1 is most likely to justify its higher price, while support lines that mostly answer questions should start from the cheaper Google model.
The clearest winners of this Gemini 3.8 Live test are companies running large volumes of simple calls, where a fast, fluid and cheap base model changes the economics. The likely losers are voice platforms that resell premium speech at a markup, since the floor price for a capable voice agent just dropped again.
For teams migrating from gemini-3.1-flash-live-preview, the switch means removing the thinking_level and enable_affective_dialog settings. Google is also pushing Gemini 3.8 Flash, which trails Claude Opus 5 on coding, and the pattern this Gemini 3.8 Live test confirms is clear: voice has become a product of its own, with its own model and its own price list.
Test Yourself on Our Gemini 3.8 Live Test
Which variant should you pick for an agent that has to complete tasks?
Show answer
Extended Thinking. It completes 68.6% of tasks on τ-Voice against 30.1% for the base model, at the same per-minute price.
What does an hour of listening cost with Gemini 3.8 Live?
Show answer
$0.84 on the base model and $3.50 on Extended Thinking, against $5.83 for GPT-Live-1 Astra.
Where does GPT-Live-1 keep the edge?
Show answer
Conversational flow: 94.9% on conversational handling, against 91.9% for Extended Thinking.
Follow the story on Horizon.
