NEWS · MODELS · #53
Google/DeepMind announce Gemini 3.8 Live and 3.8 Live Extended Thinking
Google and Google DeepMind announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking; the source material is limited to a title image referencing 'Gemini Spark' and provides no technical details. This appears to be a product/model announcement but further specifics were not included in the provided source.
KEY POINTS
- Google and Google DeepMind announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking; the source material is limited to a title image referencing 'Gemini Spark' and provides no technical details.
- This appears to be a product/model announcement but further specifics were not included in the provided source.
- An announcement of new Gemini versions signals ongoing development of Google's flagship models and is relevant to competition and capability tracking, though the source gives no technical specifics.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Google and Google DeepMind publicly introduced two new live dialogue models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, positioning them as voice-focused models with near real‑time reasoning and visual grounding.
WHY NOW
The announcements mark an incremental product/version update in Google’s flagship Gemini family emphasizing near real‑time reasoning for voice agents and enterprise use, which is relevant for tracking capability advances, benchmarking claims, and ecosystem integrations (Gemini app, Google Workspace, Search).
WHO IS AFFECTED
Developers and enterprises building voice agents and conversational apps (Google’s posts explicitly address "developers and enterprises"). Users of the Gemini app, Google Workspace, and Search are cited as product integration targets. Third‑party benchmarkers and competitors are implicated by the performance claims and cost comparisons.
CONFIRMED
— Google/DeepMind announced two models named Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking (IDs: 66, 257). — Google/DeepMind describe these as their most advanced live dialogue models with near real‑time reasoning to enable voice agents and more fluid voice conversations (66, 257). — Gemini 3.8 Live is described as built for scale and cost efficiency, with conversational intelligence, fluid dialogue, and visual grounding (66, 257). — Gemini 3.8 Live Extended Thinking is described as built for high‑complexity tasks with increased intelligence and multi‑step reasoning (66, 257). — Google/DeepMind state the models execute tools and API calls in the background while continuing the conversation, process visual inputs in near real‑time, and can detect and transition between 97 supported languages mid‑conversation (66, 257). — Google/DeepMind cite benchmark results for Extended Thinking: #1 on Artificial Analysis’ Speech to Speech Quality Index (82.6); reported agentic task completion of 68.6% on τ‑Voice and 35.1% on Sierra’s τ‑Voice‑banking benchmark; and 97.7% on Big Bench Audio (66, 257). — Google/DeepMind report Gemini 3.8 Live placed second in the Speech Agent Arena and that tests were run on the Live API on the Gemini Enterprise Agent Platform (66, 257). — A media report (The Decoder) states the models were released through the Gemini API and Google AI Studio, that sample apps are on GitHub, and reports per‑minute pricing of $0.005 for audio input and $0.018 for audio output (301).
UNCERTAIN
Missing or conflicting evidence (explicitly labeled): — Missing: official, detailed pricing and billing documentation in the Google/DeepMind excerpts (Google posts claim models are "competitive" on price but do not list per‑minute rates). The per‑minute pricing appears only in the media excerpt (301). — Missing: technical details about model architecture, latency, concurrency limits, exact API behavior, safety/guardrails, and deployment/availability rollout schedule beyond "Live API on Gemini Enterprise Agent Platform." — Potential conflict / third‑party interpretation: The Decoder compares costs and audio quality versus OpenAI’s GPT‑Live‑1 and asserts differences in naturalness and duplex behavior; Google/DeepMind excerpts do not present direct comparisons to OpenAI, so comparative quality and cost tradeoffs are not corroborated in the primary posts. — Missing: independent replication or third‑party evaluation beyond the named benchmark results; the blog posts cite benchmark scores but underlying data, evaluation methodology, and independent confirmations are not included in the supplied excerpts. — Missing/unclear: exact list and specifications for the "97 supported languages" (which languages, detection accuracy per language) and how visual input processing performs across contexts.
WHAT TO WATCH
Concrete observable signals to monitor next (what to check and where): — Official Google/DeepMind documentation and Gemini API pricing pages for confirmed billing rates and billing models (check google blog pages, API docs) (66, 257, 301). — Gemini API / Gemini Enterprise Agent Platform release notes and Live API docs for limits, latency, authentication, and availability window (66, 257). — Benchmark publications and raw evaluation data from Artificial Analysis, Sierra (τ‑Voice banking), Big Bench Audio, ServiceNow’s EVA‑Bench, and Speech Agent Arena to validate the cited scores and methodologies (66, 257). — Third‑party independent evaluations and comparisons (media, academic, or open benchmarks) that test conversational naturalness, latency, duplex behavior, and cost‑per‑hour in real applications (301). — GitHub repositories and sample apps referenced by press coverage to inspect example implementations and developer experience (301). — Product integration updates showing availability in the Gemini app, Google Workspace, and Search, and any enterprise rollout announcements or case studies from Google/DeepMind (66, 257).
WHY IT MATTERS
An announcement of new Gemini versions signals ongoing development of Google's flagship models and is relevant to competition and capability tracking, though the source gives no technical specifics.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Google and DeepMind announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.
SUPPORTEDChecked 2026-09-25 · 3 supporting
Google blog post title and excerpt present the announcement of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.
SUPPORTS Google DeepMindMEDIA · 2026-09-25DeepMind blog post title and excerpt present the same announcement.
SUPPORTS The DecoderMEDIA · 2026-09-25The Decoder reports the release of Gemini 3.8 Live and 3.8 Live Extended Thinking.
The models are live dialogue/audio models intended to enable voice agents, offering near real-time reasoning and improved multi-step/parallel reasoning for complex tasks.
SUPPORTEDChecked 2026-09-25 · 3 supporting
Google excerpt describes the two models as "live dialogue models" that bring "near real-time reasoning" and enable voice agents for complex tasks.
SUPPORTS Google DeepMindMEDIA · 2026-09-25DeepMind excerpt repeats that the models enable near real-time reasoning and are aimed at voice agents and complex tasks.
SUPPORTS The DecoderMEDIA · 2026-09-25The Decoder describes the models as audio models for developers that power voice agents and handle background API calls, visual input, and continuous conversation.
Gemini 3.8 Live Extended Thinking is reported to place first on Artificial Analysis' Speech-to-Speech Quality Index (82.6) and to lead on several agentic/voice benchmarks (68.6% on τ‑Voice, 35.1% on Sierra’s τ‑Voice-banking) and scored 97.7% on Big Bench Audio.
SUPPORTEDChecked 2026-09-25 · 3 supporting
Google excerpt lists the Artificial Analysis score (82.6), τ‑Voice and Sierra τ‑Voice-banking numbers, and Big Bench Audio score for the Extended Thinking model.
SUPPORTS Google DeepMindMEDIA · 2026-09-25DeepMind excerpt repeats those benchmark scores for Gemini 3.8 Live Extended Thinking.
SUPPORTS The DecoderMEDIA · 2026-09-25The Decoder reports that the Extended Thinking variant ranks first on Artificial Analysis' Speech-to-Speech leaderboard with 82.6.
Gemini 3.8 Live can process visual inputs in near real-time and automatically detects and transitions between 97 supported languages mid-conversation.
SUPPORTEDChecked 2026-09-25 · 3 supporting
Google excerpt states Gemini 3.8 Live processes visual inputs in near real-time and detects/transitions between 97 supported languages mid-conversation.
SUPPORTS Google DeepMindMEDIA · 2026-09-25DeepMind excerpt contains the same statements about visual processing and 97-language support.
SUPPORTS The DecoderMEDIA · 2026-09-25The Decoder notes that Gemini 3.8 Live supports over 97 languages and can process visual input.
SOURCES & TIMELINE
3Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are our most advanced live dialogue models yet. Major upgrades in intelligence and parallel reasoning make them more intuitive to collaborate with and use to execute complex tasks using your voice. Member of Technical Staff, on behalf of the Gemini Audio Team Your browser does not support the audio element. Today, we’re introducing two new models that bring adv…
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are our most advanced live dialogue models yet. Major upgrades in intelligence and parallel reasoning make them more intuitive to collaborate with and use to execute complex tasks using your voice. Member of Technical Staff, on behalf of the Gemini Audio Team Your browser does not support the audio element. Today, we’re introducing two new models that bring adv…
Google Deepmind released Gemini 3.8 Live and 3.8 Live Extended Thinking, two new audio models for developers available through the Gemini API and Google AI Studio . Gemini 3.8 Live powers voice agents that can make API calls in the background, process visual input, and keep talking at the same time. It supports over 97 languages. The Extended Thinking variant ranks first on the Artificial Analysis Speech-to-Speech Le…