• A summary of Gemini updates in January 2026 in this post here.
• Chrome has a built‑in Gemini sidebar that can read across tabs, access Gmail/YouTube/Search history, and auto-browse sites to complete tasks like booking appointments or collecting tax documents.
• Google puts Gemini directly into Google Maps so you can talk hands-free while walking or cycling, like asking it to text a friend you’re 10 minutes late or find a cafe with a bathroom on your route.
• DeepMind launched Project Genie, an AI that generates explorable 3D worlds from text, simulating physics and interactions in real time.
• Gemini’s new Personal Intelligence feature lets it securely look at your Gmail, Photos, Calendar, and other Google apps so it can do things like find your car’s license plate from a photo, suggest the right tires for your minivan, or plan a spring break trip based on your family’s past travels and interests.
• Students can take full-length free SAT practice exams in Gemini. Type “I want to take a practice SAT test” as your prompt, and Gemini will create the test.
• Google and Khan Academy are partnering to use Google’s Gemini AI models to launch new tools like Writing Coach and Reading Coach that help students improve literacy. Instead of just giving answers, Writing Coach walks students through planning, drafting, and improving their own writing. 
• Gmail is adding Gemini AI (AI inbox) so it can summarize your emails, answer questions about them, write and improve replies for you, and show a special inbox that highlights your most important messages.
• 40 most helpful AI tips from Google.
• Google’s new Gemini 3 Flash is a faster, more capable, multimodal “workhorse” model that now defaults in the Gemini app and Search, offering near–frontier-level performance at low cost for consumers, enterprises, and developers.
• Google Labs has launched Disco, an experimental AI-centric web browser. Its standout feature, GenTabs, uses Gemini 3 to turn your queries into interactive mini web apps that help you answer questions or get tasks done directly in the browser.
• Google Labs updates summary in 2025.
• Google is updating its experimental Stitch design tool with Gemini 3 to generate higher-quality UIs and adding a new “Prototypes” feature that lets you connect screens into interactive, working app flows.
• Everything Google launched in 2025 for education: 2025 Launch Guide
• Google has introduced Gemini 3, its most advanced AI model, which excels in multimodal understanding, state-of-the-art reasoning, and agentic capabilities.
• Google Maps adds “Know Before You Go” feature which gathers details about your destination before you go. It also redesigned the Explore tab, added nickname-based reviews, and now shows real-time EV charging availability.
• Google launced Google Scholar Labs, a new AI-powered feature in Google Scholar that analyzes complex research questions, searches related topics and relationships, and explains relevant scholarly papers.
• Gemini 3 Pro Image (Nano Banana Pro) offers precise studio-quality control, clear text generation, multilingual localization, and advanced creative features. Hint: Using the default “fast” model on Gemini will only kick off “Nano Banana.” To use “Nano Banana Pro,” pick the “Pro” model on Gemini.
• Upgrades to Google Photos: 1) Nano Banana AI creative image restyles. 2) Edit photos by typing or speaking requests. 3) Quick AI-generated templates for cards and headshots. 4) “Ask Photos” search expands globally. 5) New in-image “Ask” button for info or edits.
• Google launched new AI shopping features: 1) AI helps you search for products and compare them. 2) 1.Personalized and conversational shopping is made easier. 3) You can ask Google to check stores for stock. 4) Google can automatically buy items you want.
• Claims: A big challenge in image generators is said to be overcome by the updated Nano Banana 2 by Google in late 2025. Image generation tools have difficulty generating analog clock showing a specific time or wine glass filled to the brim. With the clock, the correct hand placement and readability of the clock time has been a challenge; and generating an image of a wine glass filled to the brim is difficult because most photographic data used for training only showed glasses filled about halfway. The clock and wine glass AI test is an informal challenge that evaluates AI’s ability to generate images of an analog clock showing an accurate time and a wine glass filled completely to the brim, tasks that earlier AI models struggled to do correctly due to training data limitations and conceptual understanding.

• Google has rolled out its Opal AI app builder to 160+ countries, letting teams create internal tools without coding or engineering delays.
• Google Maps is rolling out Gemini-powered features including hands-free conversational navigation, landmark-based directions, proactive traffic alerts, and AI-powered visual place discovery to make driving and exploring easier and smarter on Android and iOS in the US.
• Gemini’s Deep Research can pull in info from @Gmail, @GoogleDrive, and Chat.
• Google’s Gemini Canvas can generate complete presentations from a single prompt or uploaded notes, integrating with Workspace tools.
• Google Cloud launched Gemini Enterprise, a unified AI platform enabling businesses to deploy advanced, secure chat agents that automate workflows and integrate with enterprise apps.
• See virtual try ons: E.g., shoes.
• Google launched “Computer Use” model: The Gemini 2.5 Computer Use model lets AI systems use computers much like a human would. For example, it can click, type, and fill out forms on websites and apps. Developers can use it to automate tasks in a browser (like organizing digital sticky notes or booking appointments online)
• Google launched “Learn Your Way” which is a research experiment that uses generative AI to turn traditional textbook materials into interactive, adaptive learning experiences tailored to a student’s grade level and interests, offering tools like mind maps, audio lessons, and quizzes for personalized education. Learning is very personalized and tailored now!
• Google is upgrading its Google Photos app with Veo 3, letting users turn their still pictures into four-second video clips. Instructions: From Goole Photos app, choose “photo to video” from the dropdown menu.
• Gemini 2.5 Flash Image Maker (select Gemini Native Image, Formerly known as Nano Banana): Google’s new image maker fixes prompt precision and subject consistency, keeping faces consistent across edits like pose shifts or lighting changes, something GPT-4o and Firefly still struggle with. Try building professional headshots! Google also released image generation prompting tips. See some use cases here. Update: Gemini has overtaken ChatGPT as the No. 1 iPhone app as of September 2025 due to Nano Banana craze (e.g., turning selfies into hyperrealistic 3D figurines).
• Google Translate challenges DuoLingo: Powered by Gemini, Google Translate now offers real-time conversation support in 70+ languages and adaptive language lessons tailored for travel, study, or business—rolling out on iOS and Android.
• Google Docs adds AI-powered read-aloud feature: Google is updating Docs to let users generate audio versions of their documents using AI, making it easy to listen to content read aloud.
• Google Photos adds voice and text editing: You can now say or type edits like “brighten face” or “remove object” in Google Photos.
• Guided Learning in Gemini: Gemini challenges ChatGPT with new learning tool. Google has launched Guided Learning in Gemini, a feature like ChatGPT’s Study Mode that breaks down problems and quizzes users step-by-step.
• Coding agent Jules: Jules, Google’s AI coding assistant, is now out of beta. Powered by Gemini 2.5 Pro, it works autonomously in the cloud so you can assign tasks and let it handle bug fixes without supervision. A free version is available, with higher usage unlocked via subscription.
• Ask ChatGPT
• Genie 3: Genie 3 is a new world model that can generate interactive, consistent environments in real-time, enabling embodied AI agents to navigate and explore fantastical and realistic landscapes. The model pushes the boundaries of world simulation, with breakthroughs in environmental consistency, real-time interaction, and promptable world events.
• Storybook: You can create storybooks with a prompt like this. Storybook turns simple prompts into 10-page, voice-narrated children’s stories. Each page is illustrated in a chosen style, like claymation, comics, or anime.
• Google just launched Gemini 2.5 Deep Think, its most advanced AI yet. It uses multiple AI agents working together to solve complex questions. This new kind of multi-agent AI is powerful but expensive to run. Other labs like OpenAI and xAI are also working on similar tech to build smarter (but slower and costlier) AI.
• Google added an image-to-video tool to Veo 3 in Gemini, letting Pro and Ultra users turn photos into clips with text—advancing its multimodal platform. Here’s how to try it:
  • Open the Gemini app or gemini.google
  • Tap Video in the prompt bar (Pro/Ultra only)
  • Upload a photo
  • Enter your video description
  • Turn on sound and enjoy
• Google added three Gemini-powered AI modes to Firebase Studio:
  • Ask – Discuss and plan with Gemini
  • Agent (needs approval) – Gemini suggests changes to your app
  • Agent Auto-run – Gemini auto-applies changes to your app

• OpenAI has released GPT‑5.3 Codex, its fastest and most capable coding model.
• OpenAI launched Frontier: An enterprise platform that lets companies create AI “coworkers” that understand their business, have proper permissions, and can do real work like optimizing production, troubleshooting failures, or running sales workflows across existing systems. More info here.
• OpenAI launched Prism, a free GPT-5.2-based AI workspace that helps researchers write, edit, and analyze scientific papers using tools like LaTeX and automated citation management. Check it here.
• OpenAI will start showing clearly labeled ads at the bottom of answers for users on the free and low-cost ChatGPT Go plans (for example, travel or shopping promos after you ask about a trip), while keeping higher-tier paid plans ad‑free. Issues: This could mislead users, exploit emotional trust, and misuse sensitive personal data for targeted advertising. Update: OpenAI started testing ads here.
• OpenAI launched ChatGPT Go, an $8/month plan with expanded features and GPT-5.2 access, while confirming it will soon test ads in the Free and Go tiers for U.S. users.
• OpenAI launches ChatGPT Health, a dedicated ChatGPT experience that connects to your medical records and wellness apps so one can better understand tests, prepare for appointments, and manage health and insurance decisions without providing formal diagnoses. Right after this, Anthropic and Amazon also launched AI-health integrations.
• OpenAI now lets ChatGPT users personalize its tone and style—like making it warmer, more enthusiastic, or emoji-filled—by adjusting settings under Settings → Personalization.
• ChatGPT now has an in-chat app store, turning it into an action-oriented platform where users can access services like Expedia, Spotify, and Zillow directly within conversations to book travel, make playlists, or find homes. Developers can now build mini-apps with the OpenAI Apps SDK, submit them for review through the developer platform, and, once approved, have them listed in a new in‑ChatGPT app directory that users can discover via the tools menu or a direct apps URL.
• OpenAI released GPT-5.2-Codex, a GPT-5.2 variant built for long, agent-driven coding and cybersecurity work. It stays reliable over very long coding sessions without losing context.
• GPT-Image-1.5: OpenAI introduced a new ChatGPT Images capability, using the GPT‑Image‑1.5 model to provide more accurate, faster image generation and editing, including better text rendering and preservation of visual details.
• OpenAI introduced GPT-5.2, a large language model designed to improve reasoning and coding performance for developers and professional users. It is positioned as a direct competitor to Google’s Gemini 3 and was developed amid growing concerns about compute costs and the absence of an updated image-generation system. GPT 5.2 Pro generated a proof for Erdős Problem #397 (a long‑standing, unsolved number theory question asking whether there are infinitely many integer solutions to a specific equation involving central binomial coefficients).​
• OpenAI launched a new “shopping research” feature in ChatGPT is an interactive tool that conducts deep, personalized product research across the web—asking clarifying questions, comparing options, and then generating a tailored buyer’s guide with up-to-date details to help you choose what to buy.
• OpenAI tests group chats in ChatGPT: OpenAI is piloting a feature that lets users chat with friends and ChatGPT together in the same thread, enabling collaborative trip planning, project work, and casual group conversations.
• OpenAI has released GPT-5.1, a more conversational and adaptable version of ChatGPT featuring enhanced reasoning abilities and easier customization of tone and style for users, with rollout starting for paid subscribers and gradual expansion to all users.
• Users can now interrupt long-running queries and add new context without restarting. The model updates its response based on the new input when users click Update in the sidebar.
• “Company Knowledge:” OpenAI’s “company knowledge” lets ChatGPT connect to workplace apps like Slack, Google Drive, and GitHub so it can answer questions with accurate, business-specific context, such as automatically generating a client call briefing using recent messages, emails, and meeting notes. For example, it can summarize customer feedback by pulling insights from Slack channels, Google Slides, and support tickets, recommending next steps based on your company’s real data.​
• OpenAI launches ChatGPT Atlas, a conversational browser: ChatGPT Atlas can talk with you across tabs, remember context, and handle online tasks like booking flights or editing documents, making web browsing fully interactive.
• A New App Store: At its DevDay event (an annual developer conference hosted by OpenAI), OpenAI unveiled ChatGPT apps, monetization tools, and access to its 800M users—potentially creating the AI era’s version of the App Store. OpenAI’s 2025 DevDay revealed its ambition to turn ChatGPT into a “platform play,” integrating third-party apps so users can access services like Spotify and Zillow directly in ChatGPT, signaling their effort to become the web’s new homepage.
• Open AI’s “Agent Builder:” OpenAI is about to launch Agent Builder, a new tool that helps people create AI agents (automated programs) easily. You use a drag-and-drop interface to build workflows—like customer bots or tools for finding and comparing information—with no coding needed. It lets you connect different features, set up logic (like if/else decisions), and control how the agent works, similar to other tools like Zapier. The goal is to help anyone—especially developers and businesses—build and test AI agents quickly and with less technical skill, making it much easier to automate tasks using AI. Here are 50 sample use cases. You can access the agent builder here: https://platform.openai.com/agent-builder. Here is a guidance.
• Sora 2 video and Sora app (social media): OpenAI launched Sora 2, a model for generating hyper-realistic video, and the Sora app, a TikTok-style feed to showcase them. The app includes a “cameo” feature to insert yourself or friends into AI-generated scenes, with revocable access controls. See the difference of Sora vs Sora 2 videos here. Issue: OpenAI’s Sora Makes Disinformation Extremely Easy and Extremely Real: Examples of misinformation include ballot fraud, immigration arrests, protests, criminal acts (robbery, home intrusions), and bomb explosions and fake images of war. Takeaway: Increasing blurry lines between real and fake.
• OpenAI adds Pulse to ChatGPT: Pulse acts like a proactive personal assistant. Instead of waiting for you to ask it questions, Pulse works overnight and analyzes your interests, your recent chats, and any connected data you have with ChatGPT. Each morning, it gives you a personalized summary or set of recommendations with 5–10 AI-generated briefs from news, Gmail, Calendar, and past chats that it thinks you’ll find useful. Pro users see them as “cards,” with expansion to Plus and free tiers planned.
• OpenAI has launched Instant Checkout in ChatGPT, allowing U.S. users to directly buy products from Etsy (and soon Shopify merchants) without leaving the chat, powered by the new Agentic Commerce Protocol co-developed with Stripe. Update: PayPal has also partnered with OpenAI to enable ChatGPT users to make instant purchases using PayPal.
• Project sharing for team work and better integration: Shared projects-Teams work together, share files, and ChatGPT uses everyone’s context. Smarter connectors-Automatic integration with Gmail, Outlook, Teams, GitHub, Dropbox, and more.
• OpenAI has announced it is developing an AI-powered hiring platform called the OpenAI Jobs Platform, set to launch by mid-2026, that will directly compete with LinkedIn by using AI to match candidates with businesses and offer certifications in AI fluency through its OpenAI Academy. Related: LinkedIn is launching an AI-powered Hiring Assistant to help recruiters spot overlooked talent and focus on key tasks.
• OpenAI has added a much-requested branching feature to ChatGPT, letting users spin off new threads from any conversation to test different ideas, drafts, or code. You can find it under the ‘More actions’ menu at the bottom of any chat, then select “branch in new chat.”
• ChatGPT has the ability to monitor chats, flag suspicious activity, and report certain conversations to law enforcement. Outcome shows finding the right balance is challenging: e.g., flagging non-critical chats.
• OpenAI launched GPT-5, claiming it is faster, more accurate, and less prone to “hallucination” (80% reduction). GPT-5 automatically selects appropriate models and proactively suggests actions to accomplish tasks. The new model replaces all legacy models including GPT-4o, o3, and o4 variants because it it’s the first “unified” AI model that  combines deep reasoning and faster responses and automatically picks the right model to use for the given task. It comes with a new adaptive setup with a fast base model, a deeper “GPT-5 Thinking” engine for complex tasks, and a smart router that requires no manual selection. It can generate complex content like business plans and coding applications with minimal prompting, allows even novices to build simple software apps from short text prompts, marking a significant step towards artificial general intelligence (AGI is the level where AI possesses the cognitive abilities of a human). Related: They also released ChatGPT Prompting Guide.
• Issues with GPT-5 launch: GPT-5 aimed to advance complex reasoning with a new “Thinking” mode—but routing issues caused it to default to faster, less capable models. As a result, it underperformed GPT-4o in algebra, coding, and logic tasks, and lagged behind rivals like Claude 4.1 in real-world tests. Security flaws, including prompt injection vulnerabilities, have added to concerns over reliability. The company said they will bring back GPT-4o. CPT-5 is also not as friendly, so the updates will make it “warmer and friendlier.“
• ChatGPT now has agentic capabilities, allowing it to interact with websites, gather information, and complete complex tasks on your behalf using its own virtual computer. This unified system combines Operator’s web interaction, deep research’s synthesis skills, and ChatGPT’s intelligence, enabling it to handle workflows from start to finish. Select “Agent Mode” from the dorpdown menu when you click on the “+” sign on the chatbox.
  • Here are some interesting use cases in education:
    • A video of a student automating the homework submission using AI agent.
    • A video of an instructor automating grading of student submissions using AI agent.
• ChatGPT releases “Study Mode:” Instead of just getting quick answers, study mode in ChatGPT offers step-by-step guidance to support deeper learning, with interactive prompts, scaffolded responses, personalized support, and knowledge checks to help students build understanding. Select “Study and learn” from the dropdown menu when you click on the “+” sign on the chatbox. Sample prompt: “Act like a math tutor. Give me a calculus problem, ask me for each intermediate step (like finding the derivative, simplifying, etc.), and only reveal the next step when I get it right.“

• Anthropic launches Claude Opus 4.6, its most advanced model focused on enterprise knowledge work, with stronger planning, a 1M context window, and multi-agent “agent teams” for complex tasks.
• Claude is adding clickable mini‑apps inside so paid users can do things like send Slack messages, design in Canva or Figma, and pull Box files without leaving the chat.
• Claude launched “Claude in Excel,” an add-in that lets you chat with Claude directly inside Excel so it can understand your whole workbook, explain formulas, and safely build or edit models for you. Also, there is Claude in PowerPoints here.
• Anthropic unveils a detailed new constitution for Claude that sets out practical rules—like refusing to give self-harm instructions while offering coping resources, declining to help with malware but explaining cybersecurity basics, and answering political questions by clarifying facts without persuasion—to keep the model safe, ethical, transparent, and reliably helpful for users and developers. Long version here.
•  Anthropic released new healthcare AI updates, letting Claude connect to apps like Apple Health and Health Connect to summarize medical data and offer personalized advice.
• Anthropic’s Claude Code accelerates software development from months to days, as shown by building Cowork in just 10 days. This challenges SaaS subscriptions, with stocks like Intuit and Salesforce dropping double digits last week, while data-rich firms gain. Find 200+ Claude Code Sub-Agent Prompts here. Ideas on how to automate your life with claude code (for non-technical people). Here is a fun WSJ article about Claude Code. Fun example here. 50-good Claude Code prompts. Create presentations like this one. Top Claude Code strategies from its creator. Resources to master Claude Code here.
• Anthropic launched Cowork, a user-friendly version of its popular Claude Code for handling everyday tasks across documents, files, and the web. It is an agent that reads and edits any document on your computer. See an example of it organizing a messy desktop here. And here is a tutorial. Sample use of organizing receipts here. Update: Anthropic has added plug-in support to Claude Cowork, letting non-coders automate workflows in tools like sales CRMs, legal document systems, and marketing dashboards, with eleven open-sourced plug-ins on GitHub and easy options for creating custom ones.
• Claude has an improved Claude Code. Here is a good guide to get started with Claude Code, and here is a free course on it.
• Project Vend is an experiment in which Anthropic let an upgraded AI run a real vending-machine business, showing that it could handle basic shop tasks better than before, but it still made serious, sometimes silly mistakes when pushed or tricked by people. Watch the experiment here.
• Claude Opus 4.5 is Anthropic’s latest flagship AI model that improves coding, tool use, and long-context reasoning efficiency over earlier Opus versions, and it is now integrated across Claude’s apps and developer offerings.
• Claude released Agent Skills which let you easily add specialized instructions and code to Claude AI agents, making them smarter and better at real-world tasks by organizing skills into simple folders and files.
• Claude can now create and edit files—including Excel spreadsheets, documents, PowerPoint slides, and PDFs—directly within Claude.ai and its desktop app. Users can describe the file they need or upload data, and Claude will generate ready-to-use, professional files (with formulas, dashboards, or formatted text). This marks a shift from basic text replies to Claude completing full projects. Claude also got memory powers, can remember previous conversations and carry context across projects automatically. (A side note: GenSpark.ai has been creating similar documents for some time now-on the free version).
• Claude launched Sonnet 4.5, Anthropic’s most advanced and capable model yet, with major improvements in coding, reasoning, and safety at no added cost.
• Anthropic has released Claude for Chrome, an experimental AI tool that embeds Claude directly into the browser. It can read, summarize, and act on web pages in real time—no extra logins needed. Note: Anthropic is not the only one in the game: Perplexity’s Comet already handles browser tasks, Google’s testing Gemini in Chrome, and OpenAI is reportedly building its own AI-powered browser. Our browser experience is getting more agentic.
• Anthropic adds chat-ending feature to Claude: Claude Opus 4 and 4.1 can now end chats after repeated abusive prompts—not to protect users, but to safeguard the model. It’s part of Anthropic’s new “model welfare” effort, exploring the idea that AI might one day warrant moral consideration.
• Anthropic’s new Claude Opus 4.1 is now the top-performing AI on a major coding test, beating models from OpenAI and Google. It’s especially good at fixing bugs and working across large code files.
• Anthropic introduced “persona vectors” that let them detect and control traits like sycophancy, harmful behavior, or hallucinations in AI. This improves safety without hurting performance and reveals toxic data or subtle personality shifts that other tools miss. Similar to brain scans revealing areas linked to emotions, Anthropic mapped parts of an AI’s neural network that activate during behaviors like sycophancy, hallucination, or maliciousness—a kind of AI psychiatry. Researchers found two fixes: 1) Early detection predicts bad behavior by tracking neural activity during training. 2) Vaccine method trains the AI with harmful traits, then removes them to prevent future issues.
• Artifacts got an upgrade. You can generate apps, websites and many more fun interfaces. See samples here.
• Anthropic ended its experimental AI-generated blog: Anthropic’s AI-generated blog “Claude Explains” was shut down after a month-long pilot. The blog, which aimed to showcase Claude’s writing abilities, lacked transparency about the extent of AI-generated content. Anthropic cited the need to combine customer requests with marketing goals as the reason for the blog’s early demise.
• Claude Opus 4 and Claude Sonnet 4, set new standards for coding, advanced reasoning, and AI agents. Claude Opus 4 is said to be the world’s best coding model, while Claude Sonnet 4 delivers superior coding and reasoning. Both models can use tools during extended thinking and demonstrate improved memory capabilities.
• Claude Pro subscribers now have access to two powerful features: Integrations, which connect the chatbot to your favorite apps, and Research, which creates reports using information from the web, your Google Workspace, and other connected sources.

• Perplexity launched Model Council, a multi-model research feature that aggregates multiple models into a single answer. Instead of verifying your queries across multiple models manually, Model Council allows you to run the same query across several models at once. It queries three leading AI models, compares their answers, and delivers one unified response. See it here.
• Peplexity users can now create and edit slides, sheets, and docs directly within their prompt sessions on Perplexity.
• PayPal is giving U.S. PayPal and Venmo users early access to Comet, Perplexity’s new AI browser, along with a free year of Perplexity Pro ($200 value).
• Comet AI Browser: Perplexity has launched Comet, an AI-powered web browser that aims to compete with Chrome and Safari, leveraging its proprietary search engine to provide AI-generated summaries, links, and suggestions. Comet includes a sidebar “Comet Assistant” that reads tabs, summarizes emails, scans calendar events, and handles tasks like booking or shopping—offering seamless, cross-page support that replaces the need for multiple apps. It’s rolled out now to $200/month Max subscribers.
• Perplexity Labs: Perplexity Labs is an AI tool for Pro subscribers that helps you quickly create reports, spreadsheets, dashboards, and mini-apps by automating research, analysis, and data visualization. It goes beyond basic search by generating complete projects in minutes, leveraging tools like web browsing, code execution, and chart/image creation.
• Perplexity Pages: Perplexity Pages is a new feature that helps users easily create and share visually appealing, comprehensive articles or reports based on their research. It leverages AI to generate structured content from searches or existing Perplexity threads, and allows customization, visual enhancement, and direct sharing to a user-generated content library or other platforms.
• Perplexity Spaces: Perplexity Spaces are AI-powered workspaces for organizing research and collaborating on projects. You can use them to: (1) Organize your research: Group “Threads” (search queries) and files related to specific topics. (2) Integrate personal files: Upload and use your own documents alongside web search results. (3) Collaborate: Invite others to view or contribute to your research. (4) Customize AI: Set custom instructions for the AI within each Space. 

• Microsoft debuts MAI-Image-1, its first in-house AI image model
  The new model produces highly photorealistic images with advanced lighting and landscape rendering and already ranks in the top 10 on LMArena.
• Microsoft has added a new Agent Mode to Excel and Word in Microsoft 365, allowing users to automate the creation and review of spreadsheets and documents using AI. Through Copilot chat, users can generate presentations and detailed reports with support from Anthropic’s AI models, in addition to OpenAI’s technology. These features are currently available in the web versions of Microsoft 365 and are intended to streamline common productivity tasks by integrating advanced AI directly into familiar office applications. The Excel Agent Mode achieved 57.2% accuracy on SpreadsheetBench, ranking higher than Shortcut.ai, ChatGPT’s .xlsx agent, and Claude Files Opus 4.1, but lower than the human accuracy level of 71.3%.
• Microsoft has launched its first in-house AI models, including a fast speech generator and a new large language model for Copilot features: MAI-Voice-1and MAI-1-preview.
• Microsoft added a new COPILOT function to Excel, allowing users to summarize, categorize, and pull external data using simple formulas. Currently in beta, the feature requires a $30/month Microsoft 365 Copilot subscription. Tutorial here.

• NotebookLM launched flashcards, quizzes, and many types of reports.
• NotebookLM is rolling out a major audio upgrade with three new modes: Brief, a quick two-minute summary for fast takeaways; Critique, where dual AI hosts analyze content like in-house editors; and Debate, which presents opposing perspectives to pressure-test ideas.
• Google adds Video Overviews to NotebookLM: NotebookLM now creates visual explainers from PDFs, notes, and images, building on its audio summary feature. The update pushes Google’s AI deeper into personalized, multimodal learning — turning dense content into tailored, easy-to-digest formats.
• NotebookLM now features expert-curated notebooks on a wide range of topics, from science to literature. Users can explore content, ask questions, and share their own notebooks with the community. Some examples: Parenting advice for the digital age by psychology professor Jacqueline Nesi, longevity research from cardiologist Eric Topol, an economist report on the biggest market, tech, and political trends this year, the works of Shakespeare

Meta is buying Singapore-based AI startup Manus for $2B, integrating its agent tech into Facebook, Instagram, and WhatsApp. Manus will remain independent and cut ties with China.

Manus launched The Manus Browser Operator, a browser extension that turns a local browser into an AI-powered agent, enabling automation of tasks within authenticated platforms (like CRMs, premium research tools, and subscription sites) directly using logged-in sessions.

Manus image generation gets much better. It understands your intent, plans a solution, and knows how to effectively use image generation along with other tools to accomplish your task.

• Google’s new AI tools gives schools access to premium AI models with better data protection. Educators can share custom AI experts and students can generate personalized quizzes.
• Educators are creating custom AI experts and interactive simulations using the Gems feature in the Gemini app and will soon be able to share them.
• NotebookLM will soon also enhance the audio overviews feature with video overviews which turn your sources into engaging educational videos.

• AI Browser: Genspark AI Browser embeds AI tools directly into every webpage you visit, transforming your browser into an intelligent workspace. It provides seamless access to AI-powered search, analysis, and productivity features while maintaining robust privacy and security protections.
• AI Secretary: Genspark AI Secretary integrates with Gmail, Google Calendar, Google Drive, and Notion to automate your workflow management. This intelligent assistant handles meeting scheduling, communications, and administrative tasks to streamline your daily productivity.

OpenAI launched o3-pro, its most capable reasoning model so far. It outperforms o3 in terms clarity, instruction-following, and domain-specific tasks like business writing and programming. It is tuned for precision over speed.

ChatGPT’s Search has been upgraded to deliver unbiased product recommendations with images, prices, and links—perfect for prompts like “best espresso machine under $200.” It now includes multiple citations, trending searches, and autocomplete to make searching faster and more reliable. These features make ChatGPT a powerful, user-friendly alternative to traditional search engines.

Also, ChatGPT just made it easy to find and edit all the AI images you’ve ever generated.

Also, OpenAI is working on a screenless, wearable AI device with a camera and mic, designed with former APple designer Jony Ive. Aiming to ship 100M units by 2027, it could become a new core tech device—but faces the challenge of shifting users away from smartphones.

Also: Voice Upgrade: OpenAI just made ChatGPT’s voice sound more natural and added live translation. Paid users also get smarter responses, better memory, and new tools to pull in info from apps like Dropbox and GitHub. Free users will see more personalized replies based on recent chats. It can also translate between languages in the middle of the conversation.

OpenAI’s new “Record Mode” for ChatGPT Pro on macOS lets users record, transcribe, and summarize meetings.

Google gets a major overhaul:

Gemini & Tools Overhaul

1. It is gradually launching “Project Mariner” which is an experimental browser agent in Chrome for Gemini Ultra subscribers. It navigates, fills forms, and performs tasks via chat prompts, with strict user permission required for all actions to ensure privacy.
2. Gemini 2.5 Pro: Now includes enhanced reasoning mode, excelling in math and coding benchmarks.
3. Gemini 2.5 Flash: Now available to all users.
4. Jules (Coding Agent): New autonomous programming agent, similar to OpenAI’s Codex.
5. Stitch: New platform for designing user interfaces.
6. $250/Month Subscription Tier: Offers early access to cutting-edge features for “trailblazers and pioneers.”
7. Google AI Mode (i.e., Google search shows chatbot results on your screen) is rolling out to everyone across the US. Relatedly, Google is testing ads in AI Mode and expanding AI Overviews ads to desktop. Ads will appear within or below AI-generated answers for longer queries. This monetizes AI search while creating new advertising opportunities, likely shifting ad strategies away from traditional keyword-based approaches.

Media Models Upgraded

7. Veo 3: Generates audio with video, improved lip sync, and better real-world physics.
8. Imagen 4: Offers sharper text generation, more aspect ratio options, and 2K resolution.
9. Lyria 2: Updated music generation tools for musicians.
10. Flow: New platform to turn short clips into full films with detailed control and scene consistency.

XR, Translation, and 3D Communication

11. Next-Gen XR Glasses: Google partners with Warby Parker, Samsung, and others for AI-powered smart glasses on Android XR.
12. Beam: Makes video calls more immersive by turning people into real-time 3D avatars.
13. Live Translations in Google Meet: Near-instant, sci-fi-like translations now built into live meetings.

Google Labs – Portraits

Google Labs’ new Portraits experiment offers personalized AI coaching experiences by creating interactive AI avatars of trusted experts, starting with Kim Scott, leveraging her communication and leadership insights.

How Meta differentiates: Powered by Llama 4, Meta AI focuses on everyday consumers rather than enterprise users. Unlike competitors prioritizing productivity tools, Meta is embedding AI into daily life—across mobile, web, and even smart glasses—offering personalized, social, and entertainment-driven experiences by tapping into users’ Facebook and Instagram data.

Dark side: Meta’s AI push raises privacy concerns as its smart glasses default to recording user voices, feeding data into its systems unless users take extra steps to opt out.

1. Anthropic has expanded Claude’s capabilities with new features, including a Research tool that runs multi-source searches for detailed answers. A Google Workspace integration lets Claude access Gmail, Calendar, and Docs, with more content sources coming soon.
2. Claude added new Integrations feature: Powered by Anthropic’s Model Context Protocol (MCP), this new feature makes it much easier to connect Claude to third-party apps like Zapier, Square, and Cloudflare. Previously, setting up such integrations required significant technical expertise, but now, this new approach handles the complex setup over the web, simplifying the process for users.

• ChatGPT’s Deep Research’s new PDF export feature enables Plus, Team, and Pro users to download formatted reports, including tables, images, and clickable citations. 
• GPT-4.1 is a significant improvement over GPT-4o, with major gains in coding, instruction following, and long context understanding. They also provide a prompting guide here.

• OpenAI is also retiring GPT-4.5—its most compute-intensive model—by July, citing high costs and shifting focus to scalable models. 

• A new “Library” tab on the left menu lets users view and manage all of their AI-generated images in one place.
• OpenAI is testing a social platform to launch a new social feed. It’s not clear if it will live in ChatGPT or be a standalone app.
• OpenAI is working on o3 and o4-mini models that can achieve scientific breakthroughs. The company is considering enterprise pricing that may reach $20,000/month.
• Industry competitiveness leads to loosening safety guardrails.
• OpenAI launched its own agentic coding platform, Codex.

Below are some major updates from Google’s AI tools:

1. Google has launched Gemini 2.5 Flash, a lighter AI model that lets developers adjust how much reasoning the model performs via a token-based “thinking budget.” 
2. Gemini Advanced users can now create high-resolution, 8-second videos with Veo 2 using text prompts. Google One AI Premium subscribers also get Whisk Animate to turn images into animations.
3. Deep Research (Gemini Advanced + Gemini 2.5 Pro): Experimental Gemini 2.5 Pro model functions as a personal research assistant with strong reasoning and structured synthesis. New Feature: Audio Overviews – podcast-style narrations for portable insights. It is positioned as a full-stack tool, not just a chatbot.
4. AI-Powered 3D Remastering, reviewed 3D remastering of The Wizard of Oz for Sphere’s 16K immersive screen, demonstrates AI in next-gen cinematic experiences.
5. AI Mode Upgrade: Now includes visual recognition. Users can upload images to receive contextual insights and suggestions.
6. NotebookLM Upgrade: Integrates external sources from live web content for enhanced generative capabilities. See demo of “Discover Sources” here.
7. Geospatial reasoning: Uses satellite imagery to enhance crisis response and urban planning. See the video here.

• Pages feature: Like ChatGPT’s Canvas, Pages lets you tweak AI-generated text on the spot: highlight a passage, hit “Ask,” and tell it to adjust the tone, lengthen a section, or rephrase a thought. For now, though, it doesn’t support coding tasks.
• Microsoft Copilot can now perform online tasks like booking car rentals, concert tickets, restaurant tables or filling out forms on your behalf using simple chat prompts. This new “Actions” feature aims to automate common web-based tasks, allowing you to focus on other work while Copilot handles the details. The new Copilot can also remember details about your life, including food preferences, hobbies, and friends’ birthdays. Other features include AI-generated podcasts, vision capabilities, deep research, and customization.

• Microsoft Copilot also extended “Copilot Vision” from web to mobile. This feature turns a phone’s camera into a real-time visual search tool. You can point your camera, and chat with Copilot to discuss what you see (e.g., plant health, product specifications, book recommendations) all processed in real time.

• A year ago, Microsoft launched Phi-3 on Azure, bringing small language models (SLMs) to customers and expanding access to efficient AI tools. Microsoft’s new Phi-4 SLMs, especially the 14B Phi-4-reasoning-plus, rival much larger models like DeepSeek R1 (671B).

Meta has rolled out two new models from its Llama 4 family—Scout and Maverick—which are now integrated into Meta AI across WhatsApp, Messenger, and Instagram. Another model, Llama 4 Behemoth, is still in training, and is positioned as a future top-tier base model. Another model, Llama 4 Reasoning, is also in training. It employ a mixture-of-experts (MoE) architecture to boost efficiency and multimodal performance. Below are more details on the new models:

Scout: A lightweight model designed for efficiency, capable of running on a single Nvidia chip. It includes a 10-million-token context window and performs well compared to open-source models like Google’s Gemma 3.

Maverick: Suited for multimodal tasks that combine text and images, such as customer service bots that process photo uploads. It supports 12 languages.

Behemoth: The largest and most powerful of the models, serving as the foundation for Scout and Maverick. It shows strong performance on STEM benchmarks, exceeding GPT-4.5 and Claude 3.7, though it hasn’t been released publicly.

Typically $20 per month, the subscription includes GPT-4o—OpenAI’s most advanced AI model—featuring improved accuracy, Advanced Voice Mode, unlimited DALL-E image generation, Deep Research for literature reviews, and priority access during peak times. Students can activate the offer by verifying their enrollment via SheerID on ChatGPT’s student page.

Commentary: The focus of AI tools on students is increasing. There are multiple reasons that make sense for businesses such as: 1) students are young and they are early in their purchase history, meaning they have a potentially high Customer Lifetime Value (CLV). 2) Students are heavy users of AI for creative projects, research, and complex problem-solving. 3) Universities are switching to seeing AI as essential in education rather than ignoring or resisting the use of it.

Claude for Education is inspired by OpenAI’s ChatGPT Edu plan but with a key differentiator: “Learning Mode.” This is a new feature that leverages Socratic questioning to enhance critical thinking skills. Instead of directly providing answers, Claude responds to student inquiries using Socratic questioning, such as, “How might you tackle this issue?” or “What evidence do you have to support your conclusion?” It poses thought-provoking questions, simplifies key concepts, and provides templates for essays, research papers, and study guides rather than providing a quick answer.

The biggest issue with video generators is character and scene consistency across multiple shots. Runway’s new model solves that issue but raises concerns around copyright and job displacement in the entertainment industry.

Commentary: This is similar to the what a new published paper achieves. Here’s a short abstract of the paper: Transformers struggle with generating coherent one-minute videos due to inefficient long-range self-attention. To address this, researchers introduce Test-Time Training (TTT) layers with expressive hidden states, enabling a pre-trained Transformer to generate videos from text storyboards. Their TTT-MLP model outperforms baselines in temporal consistency, motion smoothness, and aesthetics, though artifacts remain.

Genspark launched a general-purpose Super Agent that beats Manus and OpenAI’s Deep Research on the industry-leading GAIA benchmark. It can autonomously handle multi-step projects such as creating a travel itinerary, finding restaurants, and calling them to make reservations.

Simultaneously, another startup General Agents unveiled another autonomous agent, which they call the first “real-time computer autopilot.” The agent uses a mouse and keyboard to do things like book hotels or apply for a job on its own.

GPT-4o advances image generation into a practical tool with precision and power, enabling useful and valuable image creation through a natively multimodal model capable of photorealistic outputs. The model can accurately render text, follow prompts, and leverage its knowledge base and chat context to transform images or use them as inspiration. Here are some example for different styles. And see other examples for: marketing/product ad campaigns, generating User Interface, infographic, lego filter, design-renovation ideas, visual recipe infographics, and coloring books. A viral one: the forever screenshot.

A note: Competition gets tougher by day. A day later, Ideogram released Ideogram 3.0, which it claims human testers prefer over rivals like OpenAI’s Dall-E 3 and Google’s Imagen 3. With over 4.3B style presets, it’s especially good at photorealism, text generation, and graphic design.

Another update: ChatGPT got memory boost, which lets it remember and pull information from old conversations. But the question is, should it?

The model outperforms rivals on several benchmarks, including code editing and multimodal tests. Google claims Gemini 2.5 Pro is its most intelligent model yet, with a 1 million token context window and support for longer inputs.

Google is also rolling out real-time vision to Gemini Live, letting users point their phone cameras at objects or share on-screen content for instant AI analysis. See some example uses here.

My Note: Check this platform for OpenAI Agents SDK and Responses API

Update: Google rolled out Gemini 2.5 Pro (experimental), its most advanced AI model, to all free-tier users. It is also now supporting apps, file uploads, and the newly introduced “Canvas” which is an interactive space to create and edit text documents.

Read more

xAI launched its strongest model Grok 3 with reasoning capabilities, and Grok 3 mini, a frontier model in cost-efficient reasoning. 

The Anthropic Console has been redesigned to streamline AI development with Claude. It now allows prompt sharing, extended thinking optimization, and collaboration among team members. The latest model, Claude 3.7 Sonnet, offers enhanced capabilities, including the ability to control the thinking budget.

My Comment: This sounds similar to the Boodlebox shared folder option. Or, Perplexity Spaces which are collaborative hubs designed to help users organize and manage their research. They allow you to group your threads (conversations) and files by specific topics or projects.

Read more

OpenAI recently launched their agent mode “Operator” and the research model “Deep Research.” Now, the startup released a platform that lets companies create their own AI bots for completing tasks such as customer service and financial analysis. The tool will help businesses develop their own AI agents—technologies designed to autonomously handle tasks. For instance, fintech giant Stripe tested OpenAI’s agent-building platform by creating a prototype that can analyze a sales-tracking spreadsheet commonly used by small business owners. The AI agent then generates invoices and sends them to customers. Pricing: It would charge customers based on the number of search queries, actions and data storage that the agents actually end up using, as well as for general AI model use. For instance, a company building a legal assistant agent will be charged $2.50 for every thousand queries a user makes to a knowledge base of past legal cases, according to OpenAI.

My Note: Check this platform for OpenAI Agents SDK and Responses API

Read more

Manus Mania started last week after the Chinese startup launched a new agent so impressive that some users are reportedly willing to pay up to $6,900 for beta access rather than wait for a free code. I registered for a free code but did not receive it yet but I saw many posts on Reddit from people selling their access codes. Manus AI is generating buzz just like DeepSeek did a few weeks ago.

My Note: MIT gave Manus a harsh review: It has low reliability (e.g., slow and may crash).

Read more

Most AI voices sound flat or robotic, but Sesame’s assistant is different. It feels natural, expressive, and emotionally in tune with the conversation. It comes in two distinct voices—Maya and Miles. Try it via the link below.

Read more

ChatGPT 4.5 has improved capabilities in unsupervised learning, human collaboration, and reasoning. It has broader knowledge, reduced hallucinations, and stronger aesthetic intuition and creativity compared to previous models.

My Notes: Just like with every new model release, this one with emotional intelligence didn’t keep the top spot alone. Within a few days, Alibaba launched R1-Omni that comes with emotional intelligence.

Read more

Perplexity introduced new voice mode: real time voice and real time information across several languages simultaneously.

Poe now offers Poe Apps, allowing users to build visual interfaces on top of over 100 AI models without coding or with full customization. Creators can easily develop apps like Chibify (anime art generator) and MagicErase (object remover) while leveraging Poe’s existing ecosystem. Available on the web today, with mobile support coming soon.

Read more

OpenAI has widened access to Deep Research, its AI-driven analysis engine, now available to ChatGPT Plus, Team, Education, and Enterprise users. Powered by a specialized o3 variant, it autonomously gathers, organizes, and synthesizes insights from text, images, and PDFs, offering intelligence that mirrors human analysis.

Read more

Claude 3.7 Sonnet, a model that lets users fine-tune its depth of reasoning on demand. The ‘thinking mode’ toggle shifts between instant responses and deep analysis, challenging OpenAI’s modular approach and DeepSeek’s cost-cutting strategy.

Read more

Copilot users now get free, unlimited access to Voice and Think Deeper (powered by OpenAI’s o1 model). You can have an extended conversation with Copilot using Voice and take advantage of Think Deeper’s advanced reasoning models to tackle more complex questions or tasks, anytime.

Read more

Perplexity says Comet will “reinvent” browsing, just as it did with search. This launch is part of its rapid expansion, including an AI assistant for Android, a deep research tool, and an API—all strengthening its push for an AI-first internet.

Read more

Perplexity open-sourced “R1 1776, a version of the DeepSeek-R1 model that has been post-trained to provide unbiased, accurate, and factual information” (or we can say less censored). You can access it here: https://labs.perplexity.ai/

If you haven’t had a chance to try OpenAI’s Deep Research ($200/month), you can now explore Perplexity Deep Research for free (with a daily cap). It is probably not as good as OpenAI’s version, but still an advanced research tool for free is worth checking out!

Read more

OpenAI’s ChatGPT-o1 model, known as “Strawberry,” can enhance writing by bridging human and AI-generated text, using style transfer prompts. For this, you select a sample text that embodies the style you want. Then you ask it to replicate that style.

Read more