Google AI Updates: September 16, 2026

1. Gemini 3.8 Live Took the Top Spot on Speech-to-Speech Quality

Google. Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two speech-native models that hold near real-time voice conversations, take visual input, and run tool calls in the background without breaking the dialogue. The Extended Thinking variant scored 82.6 on Artificial Analysis’ Speech to Speech Quality Index for first place, along with 68.6% on the tau-Voice agentic task benchmark and 97.7% on Big Bench Audio. Both models handle 97 languages including mid-conversation switching, and are rolling out through the Gemini API, AI Studio, Gemini Enterprise in private preview, Search Live, and Workspace. Source

2. The Pricing Undercuts GPT-Live-1 by More Than Half

Google. Gemini 3.8 Live is priced at $0.005 per minute of audio input and $0.018 per minute of output, which works out to roughly $1.38 for an hour of conversation against a floor above $3.00 for OpenAI’s comparable tier. The Extended Thinking variant leads OpenAI’s GPT-Live-1 on the Artificial Analysis speech-to-speech leaderboard, while the base 3.8 Live sits second in Speech Agent Arena. Reporting notes the tradeoff is on audio naturalness, where OpenAI’s full-duplex approach is still expected to sound better. Source

3. Google Detailed a Multilingual Stack Covering 300 Languages

Google. The company laid out the models behind its language coverage: Gemini 3.5 Live Translate handling spoken translation across 70 languages and more than 2,000 language pairs with code-switching and emotional cues, Gemini 3.5 Transcribe for noisy-environment speech-to-text, the Universal Speech Model trained on 12 million hours of audio across 100 languages, on-device TranslateGemma covering 55 languages, and a sign-language-to-text model trained on more than 50 sign languages. Supporting datasets include WAXAL for 27 Sub-Saharan African languages, Project Vaani with 30,000 hours across 109 Indian languages, and LinguaMeta, an open repository mapping over 7,000 spoken, written, and signed languages. Google puts total reach at more than 300 languages. Source

4. AlphaGenome Atlas Mapped All 9 Billion Single-Letter Human Genome Variants

Google. A companion post on scientific applications announced AlphaGenome Atlas, a map of every one of the roughly 9 billion single-letter changes in the human genome. Google also reported WeatherNext 3 delivering 50% more accurate precipitation forecasts a day or more out, Flood Hub coverage reaching 2 billion people across more than 150 countries, a breast cancer model catching 25% of interval cancers previously missed on mammograms, and a diabetic retinopathy model that has backed more than 1.15 million screenings. Source

5. ATLAS Data Shows Scientists Saving Just Under Seven Hours a Week

Google. Google opened an interactive, free-access view of its AI & Economy ATLAS dataset alongside new survey work with Google DeepMind and MIT FutureTech covering 2,600 specialized AI models and more than 600 scientists in the U.S. and U.K. Nearly half of surveyed scientists use AI daily and report time savings just under seven hours per week, though researchers flagged validation and hypothesis testing as the remaining bottlenecks. Usage mix varies sharply by country: arts and media occupations account for 19% of work-related AI usage in India, 1.6 times the global average, while computer roles account for 30% in the U.S. Source