Google AI Updates: November 6, 2026

1. Gemini Omni 1.2 Flash Reaches General Availability With 15-Second Clips and 30% Lower Video Pricing

Google released Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) on November 5, the first generally available version of its Omni video model, in the Gemini API and on Google Cloud. Clips can now be generated or edited natively up to 15 seconds, set in one-second steps from 3 to 15 seconds (Omni 1.1 stopped at 10); first-and-last-frame interpolation covers 3 to 10 seconds, and each video extension turn adds 3 to 10 seconds up to 30 seconds in total. The price is $0.07 per second of 720p video, 30% below Omni 1.1 Flash, which the API pricing lists as $12.00 per 1M video output tokens, and paid-tier rate limits are doubled. Generation holds up to six reference images more consistently and also accepts motion clips of up to three seconds and audio cues (audio references require allowlisting), editing or extending a video accepts up to two reference images alongside the input, multi-turn conversational editing works across full 15-second clips, and a standalone upscale task (task: "upscale") turns 360p drafts into 720p or 1080p and 720p clips into 1080p or 4K. The Interactions API adds a seed parameter for reproducible outputs and custom billing labels; on Gemini Enterprise Agent Platform the model runs on Standard pay-as-you-go or Provisioned Throughput at 685 tokens per second per GSU, and outputs carry a SynthID watermark. From December 3, requests to gemini-omni-1.1-flash will be routed automatically to 1.2 and billed at 1.2 rates. In Google Vids the model is rolling out from November 5 with a 3 to 15 second length slider at no additional cost and no admin control, 15-second generation is rolling out in the Gemini app to Google AI Ultra subscribers and in Flow and Google Ads, and Google says fast sports action, rapid script changes and complex physics still need work. Source Google Cloud Release notes Vids Cloud release notes

2. Gemini API Accepts Legacy Office and OpenDocument Files and Lets Tools Return Audio, Video and Documents

Google widened file handling in the Gemini API on November 5. The API now accepts PowerPoint 97-2003 (PPT), OpenDocument Presentation (ODP) and OpenDocument Spreadsheet (ODS) files and adds page-image processing for Word 97-2003 (DOC): DOC, PPT and ODP files are rendered as page images, while for ODS the model receives the data and formulas from the first sheet only, and processing of DOCX, DOTX, ODT, PPTX, XLSX and XLS is unchanged. Function responses can now carry audio, video and Office or OpenDocument files in addition to images, PDFs and plain text. In the GenerateContent API, function response parts also accept fileData, so a tool can return a Files API URI instead of inline bytes, displayName is available on both inlineData and fileData, and video parts accept mediaProcessing (STATIC or AGENTIC); in the Interactions API, function_result steps accept document, audio and video content blocks by inline data or uri, with video blocks available in v1beta and v1alpha only and accepting processing. Source

3. Google Donates Agent Substrate, gVisor and OpenRL to the CNCF and Makes Agent Substrate 1.0 Generally Available on GKE

Google Cloud is donating three open source projects that started at Google, Agent Substrate, gVisor and OpenRL, to the Cloud Native Computing Foundation after their acceptance into the CNCF Sandbox last month, and declared Agent Substrate 1.0, now generally available as a managed capability on GKE. Agent Substrate separates agent state from compute: Actors hold each agent’s memory, files and identity, Workers are reusable sandboxes drawn from a warm Kubernetes pool, and Gateways enforce inbound and outbound access and inject secrets based on Actor identity so keys stay out of the agent environment. Idle agents are snapshotted to storage and restored onto any Worker in under 500 milliseconds, and Google cites 10x the density of standard container runtimes and more than 500 suspend and resume activations per second, with Cloud Hypervisor microVMs or gVisor as the isolation boundary; the project has more than 100 contributors since public development began in May 2026, and Nous Research is building Hermes Enterprise on it. For gVisor, which runs under GKE Sandbox and Cloud Run and which Google says Anthropic and OpenAI use to contain code execution, Google says neutral governance opens a path to upstreaming host kernel patches that Linux maintainers declined while a single vendor owned the project; Ant Group, Modal and Tines have committed to join its maintainers, and it is on a path toward CNCF Incubation. OpenRL, a research preview since February, runs self-hosted RL post-training loops in plain Python on Kubernetes through a Tinker-compatible API with four primitives (sample, forward_backward, optim_step, save_state), runs existing tinker-cookbook recipes unchanged with LoRA or full fine-tuning, partitions accelerators through Kubernetes Dynamic Resource Allocation, and uses PyTorch/AutoModel for training and vLLM for sampling, with Gemma and Qwen validated and TPU support on the roadmap; paired with llm-d, it gets prefix-aware rollout routing and GPU time-slicing between sampling and training. Source

4. Google Cloud Evolves a Coding-Agent Skill With AlphaEvolve and Harbor and Finds the Winning Version Built Around Timeouts

Google Cloud published a practitioner write-up on optimizing agent skills with an optimizer built from AlphaEvolve, Harbor and Gemini 4 Argon. As a baseline, OpenCode with Gemini 3.8 Flash passed about two-thirds of its attempts on Terminal-Bench 2.0 tasks whether or not the team’s hand-written skill was loaded. The SKILL.md description and body sit between AlphaEvolve’s evolve markers in a seed program so Gemini can rewrite them, Harbor runs the agent with each candidate on real tasks in one container per trial (one GKE pod per trial, with thousands of trials in a single search), each candidate runs alongside the baseline skill at the same time so the score is the difference between them, because model latency had explained much of the score variance in early runs, and held-out tasks are kept for a final check. The post reports a 5% improvement after a few iterations; about a third of trials hit their time limit, and the best skill is a protocol built around that limit (“Ship a Working Baseline to the Target Path Early”, “Hard Stop When Green”). After about 10 hours the search scored no new candidates, because the latency check held back every evaluation while model responses were slow and full agent logs filled the controller’s disk; a skill that contained the answers added 33 points on development tasks, which the authors read as a sign that skills help when they supply specifics the agent lacks. In a small follow-up study on customer service test cases for a fictional bank, AlphaEvolve found and fixed a policy rule the team had planted as wrong and carried the fix to unseen cases, while other small studies were mixed. Source

5. Google Research Validates Cuffless Smartwatch Hypertension Screening Against 24-Hour Ambulatory Monitoring

Google Research published a study in Nature on AI-PPG-ACC-HTN, a system that flags signs of hypertension from a consumer smartwatch’s photoplethysmography (PPG) and accelerometer signals without cuff calibration. Instead of estimating point-in-time systolic and diastolic values, a 1-D convolutional encoder pretrained with self-supervised contrastive learning on 492,347 hours of unlabeled sensor data from 2,109 participants turns 15-second PPG and accelerometer windows (including the ballistocardiogram) into embeddings, which are averaged over several days, combined with the wearer’s height and passed to a regularized logistic regression that outputs a hypertension probability. In a prospective study at five U.S. sites with 196 participants who had no known cardiovascular disease and took no blood pressure medication, benchmarked against 24-hour ambulatory blood pressure monitoring, the watch reached 65.8% sensitivity, 90.0% specificity and 80.6% positive predictive value for initial screening (initial office readings: 55.3% sensitivity, 77.8% PPV). Among 48 participants with elevated office readings, it reached 78.4% sensitivity, 90.9% specificity and 96.7% PPV for confirmatory testing, against 67.6%, 63.6% and 86.2% for repeat office measurements and 89.2%, 81.8% and 94.3% for multi-day home monitoring. The work, done with Beth Israel Deaconess Medical Center and Harvard Medical School, informed the blood pressure trends Health Guardian feature in the Google Health app; Google describes the methods as exploratory and says the approach is not meant to replace arm cuffs or give a standalone diagnosis. Source Paper

6. Singapore’s GovTech Ships AntiScam Pango, the On-Device Scam Detector Built With Google DeepMind, as an Android Beta

Google DeepMind and Singapore’s Government Technology Agency (GovTech) launched AntiScam Pango on November 4 at GovTech’s STACK Conference, the app built on the on-device scam detection model the two announced in October. The beta, available for Android from the Google Play Store, runs in the background and analyzes incoming WhatsApp, Telegram and SMS messages on the phone, flags suspicious conversations as they unfold (including ones involving numbers never reported before) and explains the indicators behind each warning; it does not block or delete messages, and message content leaves the device only if the user reports it. The model was trained on scam patterns from ScamShield data, provided by GovTech’s Open Government Products unit, the police and the National Crime Prevention Council, and from an in-house classifier for suspicious links. It is Android-only for now because Android lets apps request access to incoming messaging notifications; GovTech says it is exploring integration into ScamShield, other messaging platforms and a version for iOS, with no timeline. Source

7. Gemini in Google Forms Generates Response Insights by Writing and Running Python

Google added actionable insights to Google Forms: from the Responses tab, Gemini analyzes form responses by writing and running Python code, so counts, percentages and comparisons are computed from the actual response data, then reports cross-question correlations, the main takeaways and recommended next steps. The rollout to Rapid Release domains began November 2 with up to 15 days for the feature to appear, and Scheduled Release domains follow from November 19. It is available on Business Standard and Plus, Enterprise Standard and Plus, Education Plus, consumer Google AI Pro and Ultra plans, and the Google AI Pro for Education and Teaching and Learning add-ons, and users need smart features and personalization turned on to reach Gemini in the side panel. Source

8. Google Commits $17 Billion to Its First Wyoming Data Center, Powered Partly by a 702 MW Fuel Cell Plant With Carbon Capture

Google announced its first data center in Wyoming, in Laramie County near Cheyenne, a long-term $17 billion investment developed with Tallgrass Energy, Black Hills Energy, American Electric Power (AEP) and Related Digital. AEP is contracted to build a 702 MW fuel cell development, which Google calls the largest planned of its kind, using Bloom Energy fuel cells that convert natural gas to electricity, with carbon capture and storage sending most of the resulting emissions to Tallgrass’s Eastern Wyoming Sequestration Hub; the project brings Google’s contracted carbon capture capacity to nearly 1 GW. Google is also sourcing power from gas generators operated by Tallgrass and Black Hills Energy and running a front-end engineering and design study on fitting Tallgrass’s combined cycle turbines with carbon capture. Google says it will pay for 100% of the power it uses and cover infrastructure costs driven by its growth, and the facility will be air-cooled to limit water use. Source