Gemini 3.8 Live and Extended Thinking: New Voice Models, Can Your Account Use Them?

On September 15, 2026, Google announced two voice-dialogue models across two posts, a model post and a developer post: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Google positions the two models differently: 3.8 Live is "built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding"; Extended Thinking is "built for high-complexity tasks, with increased intelligence and multi-step reasoning."

This article was fact-checked on September 18, 2026, against Google's model post, developer post, the DeepMind model card, and the Gemini API pricing page; the content reflects that day. We have not actually used either model ourselves — the capabilities, leaderboard rankings, and prices described here are Google's own claims, marked "Google says" where they are the vendor's statements; this article is not advice on using or purchasing them.

What Was Announced: Two Models, Two Positions

The model post lists several of Gemini 3.8 Live's capabilities in the same passage: it can process visual input in near real time, so the conversation has something to reference; it can automatically detect and switch languages, with the post citing 97 supported languages; and it can run tool and API calls in the background while the conversation continues, so users hear a reply first while the work finishes behind the scenes.

Extended Thinking differs in how it reasons: Google explains that this model "reasons and speaks simultaneously," and when a task needs deeper reasoning, it first bridges naturally with early verbal cues like "Let me check that…", then uses progress narration to walk the user through multi-step work happening in the background; the developer post calls the pair together "a more streamlined alternative to cascaded architectures."

The developer post separately lists key capabilities using the word "including": asynchronous function calling, visual context, alphanumeric precision (accurately parsing confirmation codes, claim numbers, and technical data), multilingual support ("97+" languages while maintaining consistent accents), and progressive content updates. The model post says voice generated by Google AI products all carries a SynthID watermark, keeping AI-generated content detectable.

Four Types of Users, and Which Model Each Can Use

The developer side is stated most clearly: Google writes "rolling out starting today" for both models, through the Gemini API and AI Studio. On the enterprise side, availability is private preview only, in Gemini Enterprise; the Customer Experience scenario, and — for Extended Thinking only — Workspace enterprise customers, are both still marked "coming soon," with no timeline given.

On the model post's availability list, the consumer-side split is: for 3.8 Live, the "For everyone" line names only Search Live; for Extended Thinking, it names Gemini Live, plus Workspace's Docs, limited to Google AI Pro and Ultra subscribers, while Gmail and Keep are open to all Google AI subscribers — so the subscription bar for Docs is higher than for Gmail and Keep. Separately, the same model post's video captions also place Extended Thinking in the Gemini App, and both the running text and a section heading name the two models together there too.

The model card's channel list is different: both models alike list Gemini API, Gemini App, AI Studio, and Vertex AI, with 3.8 Live additionally listed under Search Live, and Extended Thinking additionally listed under the Workspace that holds Gmail, Docs, and Keep. The model post's own availability list does not name Vertex AI, and does not put Gemini App under 3.8 Live either; this article presents both lists as they are, without deciding on Google's behalf which one is authoritative. The model post is also marked at the top as having been updated on September 17.

"Rolling Out Starting Today" Does Not Mean "Available Right Now"

Google's own English reads "rolling out starting today" — a start, not a completion; the enterprise side is explicitly private preview, and the scenarios marked "coming soon" carry no timeline either. As checked through September 18, 2026, the three documents, in describing availability, name only products and subscription tiers, and not one sentence states which country, region, or market is open — so this article cannot determine whether Taiwan accounts are already within the available range.

The enterprise private preview likewise states no application channel or eligibility criteria, and Google does not say when the rollout counts as complete, so users have no way to judge when they will get it.

The two posts state the language figure differently: the model post writes that 3.8 Live automatically switches among "97 supported languages" mid-conversation, while the developer post's capability list says "97+ languages." Both figures come from Google itself, and this article presents them side by side as given, without merging them into one.

Pricing and Data: One Price List for Both Models

The Gemini API pricing page lists gemini-3.8-live, gemini-3.8-live-extended-thinking, and the earlier gemini-3.1-flash-live-preview in the same pricing block, sharing the same set of prices, and it does not say the new models replace or deprecate it. The model post's line that 3.8 Live is "built for scale and cost efficiency" is a positioning statement, not a claim that its list price is lower.

On the pricing page's paid tier: input is $0.75 per 1M tokens for text, $3 for audio ($0.005/minute), and $1 for image/video; output (including thinking tokens) is $4.50 per 1M tokens for text and $12 for audio ($0.018/minute). When the developer post cites the per-minute figures, it adds a footnote explaining that they are estimates converted from the per-1M-token prices.

The pricing page also has a free tier: both input and output are listed as "free of charge," and that data is used to improve Google's products, unlike the paid tier. The quota for grounding with Google Search — real-time search verification — is listed under the paid tier: 5,000 free search requests per month, shared across the Gemini 3.x family, then $14 per 1,000 requests after that; a single request may trigger more than one query, and each one is billed.

The model card also states the specs: both models are based on Gemini 3 Pro; input covers audio, images, video, and text, with a context window of up to 128K tokens; output is audio and text, at 64K tokens. The knowledge cutoff date is January 2025, meaning what the model remembers ends there, and time-sensitive content still needs the separately billed search grounding to fill in.

What Google Has Not Said, and One Scenario Designed by the Editors

All benchmark figures come from Google's model post: Extended Thinking took the "#1 overall spot (82.6)" on Artificial Analysis's Speech to Speech Quality Index, 68.6% on τ-Voice, 35.1% on Sierra's τ-Voice-banking benchmark, and 97.7% on Big Bench Audio; 3.8 Live, meanwhile, took "second place" in the Speech Agent Arena. These are snapshots from September 15; we have not tested them ourselves, and leaderboards keep changing.

On safety, the model card states that the model Google assessed under its frontier safety framework was Gemini 3.7 Flash, and the result reached none of the "Tracked or Critical Capability Levels"; because the two new models show no meaningful new capability or material performance increase over 3.7 Flash, Google says they are "not likely" to reach those levels either — an inference from another model's results, not an assessment run on these two models themselves. The model card also notes possible hallucinations and occasional slowness or timeouts, and says work is ongoing to improve resistance to jailbreak attempts.

The following is a scenario designed by the editors, not a hands-on test result. Suppose a user is looking at a trip comparison table and asks Extended Thinking by voice to summarize the key points: by Google's account, the model would respond first, push the summarizing work to the background, and narrate its progress as it goes. That stacks three conditions that are not yet settled: the account has to be in an eligible region, and has to use the right interface and subscription tier; on top of that, the January 2025 knowledge cutoff means date-sensitive details would still depend on the separately billed search grounding.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

GPT-Live 1 Comes to the API: Can It Take Customer Service Calls, Whose Voice Is It, Will It Record?

• AI tools
• Apps & software
• AI news