What changed in AI, and what it means for your work.

Google DeepMind launched Gemini Omni 1.2 Flash and made it generally available on the Gemini API and Google Cloud. Clips go up to 15 seconds (from 10 with Omni 1.1), with multi-turn conversational edits so you can swap a product or background without regenerating from scratch. Google says it cut the price 30% from Omni 1.1, to $0.07 per second for 720p video. You can draft at 360p or 720p, edit, then upscale keepers to 1080p or 4K. Outputs carry a SynthID watermark.

Why it matters: If you ship product explainers or social ads, longer clips and cheaper follow-up edits cut rework. Check the new per-second rate against your current video stack before switching production jobs.

Google says access to Gemini, Gemini Notebook, Gemini in Classroom, and Google Search will be entirely off for K-12 student accounts until school administrators explicitly turn them on. It will share rollout details with schools in the coming months so they can prepare. Separately, Google is adding Guided Learning to Search for student accounts, coaching step by step instead of giving a final answer, starting with a pilot in select US schools later this month. Guided Learning is already available globally in Google Classroom.

Why it matters: If your kids' school or an edtech product you sell runs on Google for Education, ask the admin which of the four products they will opt into. Search is now gated with the AI suite, so 'AI off' can also mean Search off for logged-in students.

Google has published the Gemini Notebook API, a REST interface for creating notebooks in Gemini Notebook and adding sources such as web pages, YouTube videos, plain text and Markdown, along with an MCP server. Both are in an early access public preview, and Google's quickstart suggests handing a setup prompt to your coding agent. During the preview it works only with personal Google Accounts; Google Workspace accounts are not yet supported.

Why it matters: If you build research packs by hand, for example competitor pages or training videos, a script or agent can now load them into a notebook for you. Most work accounts run on Workspace, so treat this as something to prototype on a personal account for now.

Google DeepMind has released Cheminformatics Skills for Google Antigravity, installed through a Science plugin. The model handles planning, database search and reasoning, while standard chemistry tools such as RDKit, with connectors to ChEMBL and PubChem, do the actual molecular calculations. In DeepMind's own walkthrough on lung cancer drugs, the agent gathered thousands of compounds from the literature, ran analyses, built potency models and proposed new molecules with synthesis routes in about an hour, work it says used to take days. DeepMind stresses the designs are hypotheses that must still be made and tested in a lab.

Why it matters: Beyond chemistry, this is a useful pattern for any agent at work: let the model plan and explain, but have trusted, deterministic tools do the calculations where one error matters. If you work in life sciences, DeepMind is offering 10,000 scientists its top Antigravity tier free for six months, subject to selection.

Google's Gemini privacy hub says that when Keep Activity is on, your saved chats are used to improve its services, including training generative AI models, and a subset are read by human reviewers. Reviewed chats are kept for up to three years, even if you later delete your activity. The ways out are turning Keep Activity off, which keeps chats for only 72 hours and makes some connected apps such as Google Workspace unavailable, or using temporary chats. A Hacker News post flagged that paying for a personal Google AI plan does not change this; the help page makes no exception for subscribers, while work and school accounts fall under separate Workspace terms. On Hacker News (623 points), the biggest reply thread is about Pixel phones, where tasks like reminders stop working with Keep Activity off. Most other busy replies say this is expected and assume every provider does it; one commenter argued the review and retention rules are mainly about liability, and the poster replied with the page's own wording on model improvement. No real dispute on the facts. Sorted take: treat a personal Gemini account like a shared notebook, and keep client or company data in a Workspace account or a temporary chat.

Why it matters: If anyone on your team uses a personal Gemini account for work, client data in those chats can be used for training and seen by reviewers. Check the Keep Activity setting today, or move work chats to a Workspace account.

Google Threat Intelligence Group says Mandiant responded to multiple incidents in 2026 in which attackers hijacked companies' access to large language models, a tactic it calls LLMjacking, in sectors including financial services, retail, technology and transportation. In some incidents fraudulent charges exceeded $100,000 in a single day, and cumulative charges at some organisations reached $1 million before access was revoked. The most common way in was exposed or stolen cloud credentials, and some groups resell the stolen access through what Google calls LLM Mixers. Google's advice includes spending caps, per-model API quotas, secret scanning in code repositories and alerts when AI services are switched on.

Why it matters: If your team holds AI API keys or cloud accounts that can use AI services, set spending caps and budget alerts today and scan your repositories for leaked keys. Treat any unexpected switch-on of an AI service as a security event.

Google DeepMind says PlanAssist, a Gemini-based tool built with the UK government, Google Cloud and Faculty, is moving into a 12-month beta on live applications, with 12 more councils joining pilot councils Barnet, Camden and Dorset in three cohorts for 15 in total. It checks householder applications for missing documents, heritage constraints such as tree preservation orders, local and national policies and earlier objections, then drafts cited reports that officers can edit before they decide. In pilots since June, DeepMind says the average officer time on householder assessment tasks fell by 58% and more than 70% of the tool's reports were approved with no edits. Subject to successful testing, the government plans to make it available to all councils in England in 2027, with a qualified planning officer approving every assessment.

Why it matters: If your team reviews applications, claims or documents against a rulebook, this is a sound pattern to copy: AI does the completeness checks and a cited first draft, and a person signs off. Track how many drafts pass with no edits as your measure of whether it is working.

Google Cloud says spend caps in Cloud Budgets are now generally available, pausing usage of eligible services in a project once its budget is hit, without deleting resources. New since the preview are a spend caps API inside the Cloud Billing Budgets API, estimated live costs in Cost Reports (Google says caps for AI services trigger within minutes of usage), and automatic monthly resets so blocked services resume next month, with manual lifting still available. Provisioned Throughput is never halted, and default alerts go to billing administrators and project owners at 50%, 80% and 100% of the cap.

Why it matters: If your team experiments with AI APIs on Google Cloud, put a low cap on sandbox and test projects this week to limit what a looping agent or runaway script can cost. Check that the right people hold billing administrator or project owner roles so they get the alerts.

Google has opened a preview of its Connected Apps Platform, a single place for developers to build apps that work inside the Gemini app and Google Search. Developers connect a cloud-hosted MCP server, or Android MCP through the AppFunctions APIs, then register, test and submit the app in the Google Connected Console, choosing Search, Gemini or both. Once an app is live, they can track traffic, latency and error rates. Access is opening in phases, and Google says it has launched more than 50 Connected Apps across Search and Gemini, with partners such as Adobe, Experian, Localiza and Peloton already live in Gemini.

Why it matters: If you already run or plan an MCP server for your product, one build could reach customers in both Gemini and Search. Sign up for the preview at developers.google.com/connected-apps, knowing Google is adding more developers and markets over the coming months.

Google Threat Intelligence Group and Mandiant say that since late August 2026 they have seen multiple threat actors use open-source agentic pentesting frameworks such as ARTEX, PentAGI and CAIRN in real incidents, with agents carrying out significant parts of an intrusion. Google says the methods are not new, mainly compromised credentials and unpatched systems, but agents test several paths at once and work at machine speed; in one campaign tracked by Wiz, the actor typically went from a first compromised account to deep system access in around four hours. The creator of ARTEX took the original repository down in October because of abuse, but forks keep spreading. Separately, CrowdStrike reported on 29 October that VAPOR PANDA, a China-nexus group, used ARTEX, mainly with DeepSeek, against more than 200 organisations in at least 14 countries, and confirmed post-exploitation activity at nine of them.

Why it matters: These breaches still mostly start with a stolen password or an unpatched server, only faster. This week, put passkeys or MFA on admin accounts, rotate old API keys and passwords, patch anything facing the internet, and give each service account only the access it needs.

Google's Gemini API release notes for 30 October say the Computer Use tool, which lets an agent operate software through its screen, is now generally available across the Interactions and GenerateContent APIs. Google first launched a Computer Use preview in October 2025, and a June 2026 preview update added built-in browser, mobile and desktop environments, configurable safety policies and prompt injection detection. Google's Computer Use docs list Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite and Gemini 3 Flash Preview as supported models, and your own code still has to carry out each click and keystroke the model returns. The one-line release note says nothing about pricing.

Why it matters: If your team still copies data between portals that have no API, such as supplier or government sites, a generally available tool makes a browser agent easier to justify for production. Start with one read-only task in a sandboxed browser and keep a person approving anything that pays, sends or submits.

Google Research described a framework where a model alternates between normal use and an offline 'sleep' phase. During sleep it adds new low-rank expert blocks, distils what it held in short-term context into them, and trains on self-generated practice examples, aiming to learn new facts without forgetting old ones. The authors report near-zero forgetting when learning unseen languages such as Manchu and Kalamang in stages, near-perfect results on the BABILong long-context test up to 10M tokens, and 79.2% and 69.0% on AIME 24 and 25. This is a research result reported by the authors, not a product. On Hacker News (15 top-level comments when we read it) there is no consensus: some praised the clear write-up, while sceptics asked whether it holds for much larger models and objected to the neuroscience framing. Sorted take: interesting direction, nothing to change in your setup yet.

Why it matters: Today an assistant only knows what is in its training, its context window or your documents. If this holds up, models could absorb your team's new facts without retraining; until then, keep using retrieval and notes files.

Google and OpenAI both said on 30 October that they will honour a shared signal in robots.txt that lets a website tell participating AI providers, in one step, not to use its content for AI training. It builds on work by the AI Preferences (AIPREF) working group at the IETF. Google says that if a site opts out with the shared policy token and does not mention Google-Extended explicitly, it will treat that as disallowing Google-Extended, and company-specific controls keep working alongside it. Neither is live yet: Google aims to support it in 2027, and OpenAI says it aims to implement it in the coming months.

Why it matters: If you publish content online, there is nothing to change today, but this will become the simplest way to set one AI training policy for your site. Google notes that sites which specifically allow Google-Extended can appear in Gemini Apps answers and get clicks, so weigh visibility before a blanket opt-out.

Google says the Gemini app can now generate, preview and edit files without leaving the conversation: Google Docs, Sheets, slide decks, PDFs, Markdown, TXT and code. You can open a file full screen and type edits yourself, or highlight a section and ask Gemini to change it. PDFs work on web and mobile; the other formats are web only for now. It is available to all Gemini users in every language the app supports, including Workspace accounts, and there is no separate admin control for it.

Why it matters: If you draft trackers, decks or reports in Gemini, you can now fix them in place and then open them in Sheets or Slides, which cuts out the copy-paste step. Workspace admins should note this feature has no admin switch of its own.

Google says Chrome's enhanced autofill now uses Gemini with Personal Intelligence to suggest details from your Gmail while you fill in web forms, such as order numbers, tracking information and flight bookings. It is starting to roll out in the US on desktop, Android and iOS for Gemini in Chrome users who have Personal Intelligence turned on. Google says sensitive data is encrypted and that you can switch the feature off or limit what it can access in Chrome settings. Google has not said when it will reach other countries.

Why it matters: If staff use personal or work Gmail in Chrome for bookings and purchases, this will save tab switching once it reaches you, but it also means the browser reads the inbox. Decide now whether work accounts should have Personal Intelligence on, and check Chrome settings when it arrives.

Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, aimed at long-running coding work and cybersecurity. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 71.0% on CWE-bench for patching, and about half as many unsupported claims as Gemini 4 Argon in its internal testing. It is not generally available yet: a small group of cyber defenders gets access this week, and Google says developers, businesses and consumers will get it as soon as possible. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. On Hacker News (781 points, 640 comments), we read the first 25 top-level comments: 7 were confused by the naming and whether Carbon replaces Argon, 3 complained that Google announced a model people cannot use yet, 3 pointed out Carbon was not yet showing on the Artificial Analysis page Google cited, and 2 praised the price. No consensus on quality, since almost nobody had tried it. Sorted take: every number here is Google's own and the model is not open to most people, so treat this as a preview, not a reason to switch.

Why it matters: If you use Gemini for coding or security reviews, Carbon may do more for the same price as Argon once it opens up. Until then, keep what works and test Carbon on one real task when access arrives.

Google DeepMind has announced Genie 4, a world model it says can run real-time interactive simulations with native audio, including ambient sound and interactive dialogue, and can be steered with natural language. It says Gemini Omni anchors appearance and behaviour while other Gemini models help keep scenes coherent, and Gemini 3.8 Flash can act as a guiding agent that breaks goals into steps. Genie 4 is rolling out as a research preview to trusted testers first, then to Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio. DeepMind calls it early and experimental, noting drift in object permanence and other limits over longer sessions, and says outputs will carry SynthID watermarks.

Why it matters: If you prototype training environments, pre-viz or product demos, join the Project Genie waitlist when Ultra access opens, but treat Genie 4 as a draft tool until longer sessions stay stable.

Google has launched a business landing page inside the Gemini app for small business owners, covering branding, finances and customer insights. Owners can connect Google Business Profile, Google Ads and Google Workspace, plus third-party apps including Adobe, Canva, Dropbox, Shopify, Wix, Zoho and Xero. Gemini can store a brand style from a website, brand kit, fonts or colours and use Nano Banana 2.1 to generate on-brand visuals, and Google says it can turn spreadsheets, receipts and purchase records into financial graphics. The landing page, marketing and finance skills, and new connected apps are beginning to roll out globally in the coming days for all Gemini app users.

Why it matters: If you run a small shop or agency already on Xero or Shopify, try the hub once it appears in your Gemini app for one recurring job, such as weekly finance charts or on-brand social images.

Google Cloud has made generally available an Anthropic-managed option for Claude models on Google Cloud, alongside its existing Google-managed option. Google says both get new models on day zero, at the same pricing and with Google Cloud billing. With the Anthropic-managed option, Anthropic is the data processor under its own commercial terms, and customers get Claude Platform features such as Claude Managed Agents (beta), skills and code execution on day zero while still using Google Cloud IAM, Cloud Audit Logging and eligible commitments. The Google-managed option keeps Google as data processor under existing agreements, with VPC Service Controls, customer-managed encryption keys and data residency in the US, EU and APAC, but new platform features arrive only once they are integrated.

Why it matters: If your company buys Claude through Google Cloud, the choice comes down to who processes your data: Anthropic-managed gets new features first, while Google-managed keeps your existing data terms and residency controls.

Google has added a Customize menu to the Gemini app that brings custom skills and connected apps into one place. Skills hold prompt preferences for repetitive tasks, such as keeping your tone or a set response structure, and you can connect more than 40 external services as well as Google apps, with permissions you can audit or disconnect at any time. It is rolling out to all user tiers on web, mobile and the Gemini desktop app on Windows and macOS, in a gradual rollout of up to 15 days that started on 13 October 2026. Most connected apps are currently in English in the US, with select apps available in other English-speaking regions, Brazil and India.

Why it matters: If you repeat the same instructions to Gemini, save them once as a skill; Workspace admins can review and manage connected apps in the Admin console under Apps > Additional Google services > Gemini, so check that before staff start linking tools.

Google has made Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5 available in Google Antigravity, its agentic coding tool, for Google Cloud customers, alongside Gemini models such as Gemini 3.8 Flash and Gemini 4 Argon. An administrator must enable them first; they then work across Antigravity 2.0, the Antigravity CLI and the Antigravity IDE extensions for VS Code, JetBrains IntelliJ and Visual Studio. Claude usage does not draw from the pooled Gemini Enterprise monthly token quota and is billed pay as you go through Google Cloud Billing, so overages must be enabled. Google suggests having one model family write code and another review it, and using cheaper Flash subagents for routine lookups.

Why it matters: If your team already codes in Antigravity, you can now have one model write and a different one review inside the same tool; set spend limits first, because Claude usage is billed separately.

Google Research has publicly released Deep Review, an agentic AI system that checks long technical documents for logical consistency, verifies mathematical statements, grounds facts and compares the work with existing literature. It grew out of the Paper Assistant Tool, which gave opt-in, pre-submission feedback on more than 16,000 manuscripts at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, and Google says it now covers documents in fields such as law and finance. Every user gets two free reviews a week, and heavier users can connect a Gemini API key for higher limits and API access. It accepts PDFs and plain text for now. Google says uploaded documents are never used to train its models and are stored for one year only so they can be shown in the Deep Review interface. Sorted take: a cheap second reader for long reports, contracts or papers, but treat its findings as prompts to check, not a sign-off.

Why it matters: If you write or review long technical, legal or financial documents, try two free reviews this week on a draft you already know well, to see what it catches and what it misses.

Google DeepMind and GovTech Singapore fine-tuned a small Gemma model to spot scams in chat conversations, running entirely on the phone. It works in two steps: a lightweight classifier screens incoming messages, then the Gemma model reads more of the conversation to confirm a scam or dismiss a false alarm, and shows a warning before the user shares money or details. Training used anonymised data from scam reports and intelligence feeds, with no personal data, inside a government-managed cloud environment. Scams make up over half of Singapore's reported crimes, with more than S$4 billion lost since 2019. It is not in apps yet: rollout will be phased, needs user consent, and Google says neither it nor the government can read the chats. No accuracy figures were published.

Why it matters: If you run a bank, wallet, marketplace or support team, this is a template worth watching: a small local model trained on local scam patterns that warns users without sending their chats to a server.

Google DeepMind said on 13 October 2026 that Gemini 4 Argon is now available worldwide for developers and enterprises through the Gemini API in AI Studio and the Gemini platform, for Google AI Ultra subscribers in Antigravity, and for Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Google prices Argon at an introductory $2 per million input tokens and $10 per million output tokens (cached input 95% off), rising to $4 / $20 after 30 November 2026. The model has a 1M unified input-output token limit via the API. Google claims Argon ranks first on the Vals Index across finance, coding, legal and tax work, and shows internal wins such as more than 35% more agentic code submissions with nearly a third fewer rollbacks in Antigravity, plus speed-ups on ADK session prep, xTB chemistry, and Gemma inference. Safety notes include a misalignment monitor and 0.7% attack success on Gray Swan indirect prompt injection. Figures are Google's own. On Hacker News (about 943 points, ~730 comments), many praise Argon's clear writing and depth for non-coding work, while a large camp is angry at the rollout: "Try Gemini 4" buttons that still open older models, Pro vs Ultra confusion, and slow access. Coding comparisons split between people who find it competitive and people who still prefer Claude Opus 5.5. No consensus. Sorted take: treat this as a real frontier option for paid Google accounts, but verify you actually have Argon in the model picker before you change workflows.

Why it matters: If you already pay for Gemini Pro or Ultra, or use the Gemini API, Argon is worth a side-by-side on your hardest docs and coding tasks this week, especially while the intro price lasts through November.

Google published a developers post on 13 October 2026 on using Gemini 4 Argon across the robotics stack. Argon scores 87% on Embodied Reasoning (ERQA), which Google says beats other frontier models on pure reasoning for physical 3D problems. Demos include Boston Dynamics Spot learning to open a foot-pedal bin after practising in MuJoCo, zero-shot arm control with visual grids, closed-loop steering of the Gemini Omni world model, and Real2Sim twins from robot cameras, egocentric human video, or paper sketches. Another demo claims Argon raised a VLA success rate from 56% to 75% and cut completion time from 35s to 24s by synthesising 4,500+ sim episodes from one photo. Figures are Google's own demos and benchmarks. Access is the same Argon rollout: AI Studio and Gemini API (`gemini-4-argon`) on the paid tier, and Antigravity for Ultra subscribers. On Hacker News (about 184 points, 45 comments), the largest thread asks how much more efficient models and chips must get before Argon-class reasoning is practical on a robot's battery; replies split between "years away" and "run it over the network from a nearby data centre". A smaller group are hobbyists building their own robots, including a snow-sweeping prototype and a Unitree humanoid controlled by an LLM, who are encouraged by zero-shot control. One sceptic says it falls short of Figure 3 and Unitree's agile robots. No consensus. Sorted take: impressive demos for robotics teams already on Google's stack; not a plug-and-play product for most operators yet.

Why it matters: If you build robot or simulation workflows, Argon's video-to-sim and promptable control loops are worth a spike once you have API access, but treat the success-rate claims as lab demos until you reproduce them.