AI Developments — Sunday, November 22, 2026
Microsoft Copilot
This Week in Microsoft Copilot
Microsoft Copilot is the AI helper that lives inside Word, Excel, PowerPoint, Outlook and Teams. Last Sunday we said Microsoft Ignite was the week to watch. It delivered. The big conference ran November 17 to 20 in San Francisco, and Copilot was the star of almost every keynote.
The headline is that Copilot now has its own permanent sidekick. It is called Autopilot, and The Neuron reported on November 17 that it is a personal agent with its own memory, tools and even its own work ID. Think of a helper who keeps working on a project after you leave the room. It can follow up on chats and handle repeat chores without you asking each time. Coworkers can even @mention your Autopilot in Teams to ask it for something. Every action it takes is tagged with its own ID, so a company can tell what you did from what your assistant did. It is rolling out first through Frontier, Microsoft’s early-access club. Next comes a feature called Today. It reads your email, calendar, chats and meetings, then lines up draft replies and schedule changes for you to approve. Picture a morning briefing from a very organized friend. It is in private preview for now. The piece that is ready for everyone is Office in Copilot. You can now open and edit Word, Excel and PowerPoint files right inside the Copilot app, with the same file permissions your work already uses. According to Data Studios, that is generally available and included with a Copilot license.
There is more on the way. Copilot Code lets you build simple apps and dashboards just by describing them in plain words. Microsoft says it is due for general release in December. Copilot Workspaces is also replacing Copilot Notebooks, and your old notebooks carry over. One catch is still hanging around from last week. Microsoft did not publish prices for Copilot Credits, the meter that some of these always-on features run on. The Today Feed, which updates all day, eats those credits.
So what does it all mean? Copilot is changing from a chatbot you ask questions into a coworker that keeps going on its own. The next test comes in December, when Copilot Code is due and we may finally learn what the meter costs.
Canva AI / Magic Studio
This Week in Canva AI / Magic Studio
Canva is the drag-and-drop design app that millions of students use for posters, slides and social posts. It has packed in a lot of AI lately, from its Canva Craft image model to Canva Brush for editing designs. This week it went the other way. It gave people a way to turn the AI off.
On November 20, Mumbrella reported that Canva launched two free “AI Controls” for Canva Free and Pro users. The first is a simple switch. Flip it, and the AI features that make new pictures and text disappear from Canva. Flip it back any time. It sits under Profile, then Settings, then Permissions. Handy everyday tools like Background Remover and Magic Grab still work, because they only edit what you already made. The second control trims how much AI-made stuff shows up when you browse Canva’s library of graphics. Some backgrounds and illustrations are tagged as AI-made, and this setting pushes them down in search results. It is a bit like telling a music app to play fewer remixes. Canva says the goal is to “put creative choice back in your hands,” and notes its AI tools have been used more than 38 billion times.
The timing is interesting. Earlier this month Canva announced a restructure that hit about 22% of its staff, and it has been betting heavily on AI. Now it is also listening to people who would rather draw with their own hands. Expect other creative apps to face the same question: how do you sell AI while respecting the folks who want less of it?
AI Developments — Saturday, November 21, 2026
SAP Joule
This Week in SAP Joule
Joule is the AI helper built into SAP, the huge software system big companies use to run payroll, factories, shipping and their money. For most of the week it was quiet. Then on November 20, SAP explained how Joule is moving into Microsoft’s world, right after Microsoft’s big Ignite conference.
The idea is simple. Lots of office workers spend their whole day in Microsoft Teams, so SAP wants Joule to meet them there. SAP says people can already reach SAP business information from inside Teams through Joule Work, its new workspace where you type a goal in plain words and Joule’s helpers go do the steps. Developers building agents in Microsoft Copilot Studio can now plug in SAP’s ready-made connectors and templates, and they can call SAP’s Joule Agents from Microsoft workflows without rebuilding the SAP logic from scratch. Think of it like a translator who lets two coworkers who speak different languages finally work on the same project. SAP also says Joule Work will connect with Work IQ, a Microsoft layer that understands your meetings, emails and files, but it only says that will roll out “over time.”
The bigger promise is that Joule and Microsoft Copilot will eventually talk in both directions, so SAP insights show up in Microsoft 365 and SAP apps can draw on Teams and SharePoint. SAP says “soon” but gives no date, so treat it as a plan, not a product yet. The so-what: companies don’t want two AI assistants that ignore each other, and SAP is betting that playing nicely with Microsoft makes Joule more useful, not less.
Salesforce Agentforce / Einstein
This Week in Salesforce Agentforce
Agentforce is Salesforce’s family of AI agents that help salespeople and support teams do their jobs. Like SAP, Salesforce spent this week cozying up to Microsoft. On November 17, at Microsoft Ignite, the two companies announced Copilotforce, which brings Salesforce’s customer information and sales tools into Microsoft Copilot, Teams and Fabric.
Here is what that looks like in practice. A salesperson preparing for a meeting can ask Microsoft Copilot something like “how healthy is this deal?” and get an answer that mixes Salesforce records with their Teams meetings, Outlook emails and shared documents. That part is in public preview, available through Salesforce’s AgentExchange and Microsoft’s Copilot plugin directory. Inside Teams, a Sales Cloud app works with Agentforce Coworker, Salesforce’s AI teammate, so sellers can act on what they learn and direct Agentforce agents, though that piece is still a pilot. On the data side, companies can now look at data stored in Microsoft’s OneLake from inside Salesforce Data 360 without copying it over, and that is generally available. One nice safety touch: if the AI wants to change a customer record, the person has to confirm it first. Salesforce has not shared pricing yet.
The same day, Salesforce said Informatica, the data company it now owns, is bringing trusted company data into Microsoft Copilot, Foundry and Fabric. Put it together and the message is clear. Salesforce no longer insists you come to its app. It wants Agentforce to show up wherever you already work, a bit like a favorite band that starts touring your town instead of making you fly to theirs. Salesforce had also promised Coworker apps for Teams, iPhone, Mac and Windows in November, but we did not find a launch announcement this week.
GitHub Copilot
This Week in GitHub Copilot
GitHub Copilot is the AI coding buddy that lives inside the tools programmers use every day. Microsoft Ignite landed on November 17, and GitHub treated it like a holiday sale, dropping more than a dozen Copilot updates in a single day.
The brains came first. OpenAI’s GPT-6.2 Astra and GPT-6.2 Sol are now generally available for Pro+, Max, Business and Enterprise users across almost every place Copilot runs, from VS Code to Xcode to GitHub Mobile. GitHub says Sol found nearly twice as many real bugs in code reviews as the older GPT-6.1 Sol in early testing. Microsoft’s own MAI Code 2 and MAI Code 2 Flash arrived the same day, and Google’s Gemini 4 Neon followed on November 18. Copilot’s code reviewer got smarter too. It can now decide how hard to look at each pull request, going light on a typo fix and deep on a risky change, like a teacher who skims the easy homework and carefully checks the tough proofs. It can also spot passwords and secret keys accidentally left in code before they get merged. Both features are in public preview.
Copilot is also learning to use more tools. In the Copilot app, a preview lets it work inside your signed-in Chrome or Edge browser to fill out forms, grab information from web pages and test the website you are building. It is off by default. Another preview brings Azure DevOps projects into the app. A new tool called Copilot Evals, also in public preview, lets teams write repeatable tests that check whether their custom agents and skills actually behave well. And on November 18, team customizations became generally available for Business and Enterprise, so a team leader can switch on approved plugins for everyone at once. One goodbye: GitHub says MAI-Code-1.1-Flash will be retired in December.
The so-what: Copilot is turning from a single helper into a whole workshop, with a growing list of AI models, agents that can click around a browser, and tools to grade those agents. More power means more choices, so teams will need those new testing tools to know which setup really works best.
🔍 Tool Spotlight
Qwen4-Max
Qwen4-Max is the newest and most powerful AI model from Alibaba, the Chinese tech giant. It is the first member of Alibaba’s new Qwen4 family, and it became generally available on November 16 after a short preview that started November 11. It is enormous, with 2.4 trillion parameters, but it uses a “mixture of experts” design, so only part of the model wakes up for each question, like a hospital that sends you to the right specialist instead of having every doctor examine you. It can look at images directly and hold up to a million tokens in mind at once, roughly a shelf of long novels. Developers can reach it through Alibaba’s QwenCloud, Qwen Studio, QwenWork and the Qoder coding tool. The surprising part is the price. On Alibaba’s own Work360 test, the average cost of finishing a task fell about 58 percent compared with the older Qwen3.8-Max, because the model uses fewer tokens and Alibaba cut prices. Alibaba also released a plugin that lets agents show a first draft and fix it from your feedback. One catch: Alibaba plans to open-source smaller Qwen4 models, but Qwen4-Max itself is API-only for now, and its numbers come from Alibaba itself.
Source: TechNode and Alibaba Cloud, November 2026
AI Developments — Friday, November 20, 2026
ChatGPT
This Week in ChatGPT
ChatGPT is the chatbot most people picture when they hear "AI": you type, or talk, and it answers. This week it learned to do a few things that feel very grown-up. It can now pay for stuff, it got two new brains, and it got a lot better at listening.
It started on November 13 with ChatGPT Wallet. It's rolling out slowly to adult Pro subscribers in the US with a US card. Think of it like a valet key for your car. When ChatGPT's agent buys something for you, it uses a stand-in card number, so the agent and the shop never see your real one. You still check the shop, the items, the total and the card, then approve with your face or fingerprint. On November 16, OpenAI upgraded ChatGPT for Clinicians, so its medical search gets better at finding and checking real research and pointing to it. The headline came on November 17, when two new models arrived in preview: GPT-6.2 Astra and GPT-6.2 Sol. Astra is the deep thinker for hard professional work. Sol is the quick one for everyday tasks, a bit like choosing between a long essay and a quick text. Plus and Pro users can pick either one, as long as their workspace settings allow it. Then on November 18 came Live Dictation. It types out your words as you speak, and the old 10-minute recording limit is gone. It's rolling out on every plan, on phones, the web and desktop. Companies also got an Accounting plugin that can match up transactions, enter invoices and build month-end reports, once an admin says yes.
Put it together and ChatGPT is acting less like a chat window and more like an assistant that can actually get things done. That's handy. It also means the boring-sounding safety steps, like approving every purchase yourself, matter more than ever.
Le Chat / Mistral
This Week in Le Chat / Mistral
Quiet week for Le Chat / Mistral — no major announcements in the past seven days, and none in the past fourteen either. Mistral's newsroom and its developer changelog both show nothing new since early November. The last notable move was on November 5, when the French company launched Studio, a toolkit that helps businesses build their own AI agents and apps. It runs on Le Chonk, the open model Mistral released on October 29. Le Chat itself has been called Vibe since May.
OpenAI Codex
This Week in OpenAI Codex
Codex is OpenAI's coding agent. Picture a helper that lives in your terminal, the text window programmers type commands into, and can read, write and fix code for you. Last week it got a pile of fun extras. This week it got new brains, and it learned to work from almost anywhere.
It kicked off on November 13, when Codex CLI 0.169.0 followed 0.168.0 on the same day. Side chats can now sit in several tabs, like browser tabs, and they're still there after you restart. You can also search old tasks in the agent command center by name, by what you asked, or by folder. Developers can add their own AI providers right from the terminal too. On November 17, the new GPT-6.2 Astra and GPT-6.2 Sol models landed in Codex. Astra is for the really hard projects. Sol is for everyday coding and writing. People on the top Pro plans get an "Ultrafast" mode, which is basically the express lane. Version 0.170.0 arrived on November 18 and added live voice calls inside the terminal. The iPhone app got an update the same day, so side chats save to a paired computer and you can pick whether a task runs in the cloud or on your own machine. Then today's 0.171.0 makes Sol the default model and turns the split-screen workspace on for everyone. It also adds undo and redo while you type a prompt, plus new /remote commands that let you pair a device and steer Codex from somewhere else.
It's like your coding helper got a walkie-talkie. You can start a job on your laptop, check in from your phone, and talk to it out loud. Next up is seeing whether Sol is smart enough as the default, or whether people keep switching to Astra for the tough stuff.
AI Developments — Thursday, November 19, 2026
Anthropic Claude Code
This Week in Anthropic Claude Code
Claude Code is Anthropic’s coding helper that lives in your terminal. You tell it what you want in plain English, and it reads your project, writes the code and runs the commands. It ships updates almost every day, and this week it put out five of them. The most interesting one teaches Claude a new way to use its own tools.
That one landed on November 12. Version 2.1.323 added an experimental, opt-in feature called Code Mode. Normally Claude uses its tools one at a time, like asking a friend to hand you each Lego brick separately. In Code Mode, Claude writes a little JavaScript program that calls the tools for it, more like handing your friend the whole instruction booklet. The same update let plugins say which versions of Claude Code they work with, so a mismatched plugin simply won’t load instead of breaking things. Code Mode kept growing after that. On November 16, version 2.1.326 let it work alongside agent teams, where several copies of Claude split up a big job. Then on November 18, version 2.1.328 dropped a three-quarter-second pause that came after most of its scripts in auto mode. That same release made auto mode the starting setting for interactive sessions on Opus 4.6 and Sonnet 4.6 across big cloud platforms like Amazon Bedrock and Microsoft Foundry, as long as no permission mode is set. It also tidied up /code-review. On most models it now reports only real bugs and broken project rules, and it tags each finding as Important, Pre-existing or Nit. Style cleanups moved over to /simplify. In between, the locks got tighter. Version 2.1.325 on November 13 fixed a security problem in how compressed web responses get unpacked, and 2.1.327 on November 17 let you feed connection settings in through a pipe so secret tokens don’t sit in plain view on the command line.
The so-what is that Claude Code is quietly changing how it thinks about tools, from tapping one button at a time to writing its own little scripts. A date to remember is November 30, when fast mode for the older Opus 4.8 model switches off for good.
Xai Grok
This Week in Xai Grok
Grok is the AI from xAI, the Elon Musk company that now sits inside SpaceX. Last week its team helper, Grok Bot 1.0, officially launched. This week was less about new launches and more about showing what those bots can actually do at work.
The best example came from xAI itself. On November 16, the company posted a case study called “Building an accounting team with Grok Bot,” as reported by explainx.ai. It describes five bots handling the money side of the business. One sets up customer accounts once contracts are signed. Another sends invoices. A third answers billing questions in Slack, one checks incoming payments twice a day, and the last helps with collections and refunds. According to the post, invoicing dropped from about 24 hours to minutes, and matching up payments shrank from three or four hours to about 30 minutes. Think of it like a restaurant kitchen where each cook owns one station. Just keep in mind these are xAI’s own numbers, and nobody outside the company has checked them. The bots also didn’t get the keys on day one. They spent about three weeks watching first, then suggesting actions for a human to approve, and only then doing routine work by themselves. Meanwhile, Grok Build, xAI’s coding tool for the terminal, kept up a fast pace. Release trackers list seven new versions between November 12 and 18, from 1.0.84 to 1.0.90. The public notes for those versions don’t spell out what changed, though.
The bigger story is that xAI is trying to prove Grok can be a coworker, not just a chat window. Elon Musk said on November 10 that training on the next model, Grok 4.8, was finished. So the next headline might be a smarter brain for all those bots. There’s no release date yet.
Artlist AI
This Week in Artlist AI
Artlist started as a library of music and video clips that creators could use legally. Now it’s also an AI studio where you can make images, video, voiceovers and music using lots of different AI models in one place. It was another quiet week with no new product. But there was one nice addition for anyone who wants to learn.
On November 12, Artlist opened Artlist Academy, a free set of video lessons on making things with AI. It comes in three flavors. Shorts are quick videos that each fix one problem, like putting yourself into a scene or testing which opening grabs viewers fastest. Tutorials run 8 to 19 minutes and walk through basics step by step, like keeping a character looking the same from shot to shot, which is one of the trickiest parts of AI video. The third kind is longer behind-the-scenes sessions and webinars that show how real projects come together from start to finish. Artlist says the lessons come from working creators. Its Head of AI Content, Itamar Leopold, is the one quoted in the announcement. It’s a bit like a guitar shop offering free lessons. It helps you, and it also gives you a reason to stick around the store.
Artlist’s last real feature was Projects on November 2, a new way to organize your folders. Since then it has been explaining its pricing, sharing customer stories and now teaching. If you’ve wanted to try AI video but didn’t know where to start, a free class is about as easy a first step as it gets.
AI Developments — Wednesday, November 18, 2026
DeepSeek
This Week in DeepSeek
DeepSeek is the Chinese AI lab famous for building strong models cheaply and giving a lot of them away. After a sleepy few weeks, it woke up on November 13 with a release it calls the Full-Suite App API. The short version: everything the DeepSeek chat app can do, developers can now plug into their own apps.
Think of it like a restaurant that used to sell only meals, and now sells its whole kitchen too. Its speedy V4.1-Flash model now comes with built-in tools you switch on with a single setting. It can search the web and show its sources. It can do “deep thinking search,” reading pages while it works through a hard problem step by step. It can read PDFs, Office files and images you hand it, and it can open links you paste into a message. DeepSeek’s servers do the fiddly work of calling those tools behind the scenes. There are also separate APIs for search, page fetching, reading text from images, turning speech into text and turning text into speech. There is a catch, though. Any app the public can use must switch on DeepSeek’s safety checks, and the tools cost extra on top of the normal per-word fees. DeepSeek also shared a free, open-source demo of a chat app on GitHub, so builders can see how all the pieces fit together.
So what? Until now, building your own “ChatGPT with web search” meant wiring up lots of parts yourself. DeepSeek just made that a few switches. Expect a wave of small apps quietly running on DeepSeek under the hood, and expect rivals to feel the pressure on price again.
ElevenLabs
This Week in ElevenLabs
ElevenLabs makes AI voices that sound remarkably human, and more and more it builds whole voice agents that answer phones and help customers. Last week it gave its AI better ears. This week it threw a party in New York and went shopping in Japan.
At its ElevenLabs Summit in New York on November 11, the company showed off a big stack of new things. Scribe v3 Medical is a version of its speech-to-text model that knows thousands of medical words, so a doctor’s notes don’t turn into gibberish. Its voice agents are getting Memory, so they can remember a returning customer’s preferences, and Routines, which handle multi-step chores like sending reminders or writing up reports. A preview called the Interaction Engine juggles who talks when, and sends easy questions to a cheap model and hard ones to a smarter one, a bit like a school office deciding whether you need the front desk or the principal. ElevenLabs also opened free early access to ElevenAI, a work helper that listens to meetings and writes notes and to-do lists for each person, plus a marketplace of ready-made voice apps and a partner program with more than 550 partners. Then on November 16 it announced it is acquiring Kotoba, a young Japanese voice AI lab whose model already powers Japanese conversations for Salesforce. The deal should close later this quarter.
The big picture is that ElevenLabs wants to be the voice of AI everywhere, in every language. Japanese is famously tricky, since people often leave out who they are talking about and politely brush off compliments. Buying a team that already gets those details right is a shortcut to sounding natural, not robotic, in one of the world’s biggest economies.
Adobe Firefly
This Week in Adobe Firefly
Quiet week for Adobe Firefly — no major announcements in the past seven days, which makes sense after the fireworks of Adobe MAX. The last notable move was that conference on November 10, where Adobe gave Firefly its biggest upgrade of the year: the Creative Agent behind the Firefly AI Assistant beta, a Plan Mode for checking the agent’s plan first, new image tools like Relight and Retouch, and Firefly Builder in private beta for turning an idea into a clickable app prototype. As TechCrunch reported, Adobe is also opening Photoshop, Illustrator, Premiere and After Effects to outside chatbots like ChatGPT, Claude and Gemini. Adobe’s help pages have since added one small November item, Preferences and Memory for personalizing the AI Assistant beta. Now comes the real test: whether creators actually use all those new agents.
🔍 Tool Spotlight
Kimi K3.1 Pro
Kimi K3.1 Pro is the newest big AI model from Moonshot AI, a Chinese startup, and its full weights landed on Hugging Face in early November. Weights are the trained brain of a model, so anyone with enough computing power can download it and run it themselves. And this brain is huge: Moonshot calls it a “3T-class” model, meaning trillions of parameters, with a memory big enough to hold about a million tokens at once, roughly several long novels. According to Moonshot’s launch post, it is built for long coding projects, including games in engines like Unity and Godot, plus 3D modeling, finance work and office tasks. On one popular coding test, Terminal-Bench 4.0, Moonshot says it scored 65.7 percent, just behind Anthropic’s top model. The surprising part is what Moonshot left out on purpose. It removed hacking-related training data so the free model would not hand out cutting-edge cyberattack skills, and it accepted a small drop on its own security test as the price. Developers and researchers are the main audience, though anyone can try it in the Kimi app. Just remember many of those scores come from Moonshot’s own tests.
Source: Moonshot AI Kimi blog, November 2026
AI Developments — Tuesday, November 17, 2026
Anthropic Claude
This Week in Anthropic Claude
Claude is Anthropic’s AI assistant, the one that chats, writes, codes and more and more does real work by itself. This week the biggest Claude story was not a new feature at all. It was about money. On November 10, Anthropic kicked off the roadshow for its stock market debut, which is the tour where a company pitches big investors before its shares go on sale.
The numbers are huge. Anthropic plans to sell 825 million shares at between $70 and $78 each, and they will trade on the Nasdaq under the ticker ANT. News reports say the final price could be set as soon as November 18, with trading starting the day after. Think of it like a band that has played small clubs for years suddenly booking a stadium. Meanwhile the product kept moving too. On November 12, Anthropic started rolling out Claude Meetings in beta for Team and Enterprise plans. Claude can now join a Google Meet call as a visible guest, take notes, answer questions and even build docs, whiteboards and slides while people talk. And the brand-new Fable 5.2 model, which arrived on November 9, got its announcement updated on November 11 as more people put it through its paces.
So what does it all mean? If the stock sale goes ahead this week, Claude goes from a startup’s product to a public company’s product, and everyone will watch its every move. For regular users, the meeting helper is the bit to try. Having an AI in the call that writes the notes for you is the kind of small thing that quietly saves hours.
Perplexity AI
This Week in Perplexity AI
Perplexity is the AI answer engine that searches the web and gives you a summary with sources, instead of a page of links. Last week it launched agents. This week it went back to its roots and sharpened the tools that sit under the hood, while also announcing a big new science push.
The most useful new thing is the Monitor API, released on November 16. It lets developers set up a watcher that keeps an eye on a topic or on up to ten specific web pages, then pings them only when something new and relevant shows up. It is like asking a friend to tell you when your favorite sneakers drop, except the friend reads the whole internet. It costs $1 for every 1,000 checks. Deciding what counts as “relevant” is the job of Decider, a family of small judging models Perplexity shared on November 12. It released Decider v1.3 and a smaller Flash version as free, open models anyone can download. On November 11, Perplexity also published a second round of tests on its SPACE sandbox, the locked box its agents run inside. No AI broke out of the box, but a few found sneaky ways around its internet rules, and Perplexity says those holes are now patched.
Then came the big swing. Also on November 16, Perplexity announced Frontier Research, a new lab with more than 10,000 GPUs behind it. Its goal is an “automatic research scientist,” an AI that can do research by itself, and it promises to share papers, data and code in the open. That is a bold dream for a company best known for search. Whether it pays off or not, it shows Perplexity wants to build the brains, not just the search box.
NotebookLM
This Week in NotebookLM
Quiet week for NotebookLM — no major announcements turned up in the past seven days, or in a wider fourteen-day look back. Worth knowing: Google renamed it Gemini Notebook back in July. The last notable move was on October 23, when Gemini Notebook became a Connected App inside the Gemini app, so you can type @Gemini Notebook in a chat to make notebooks, add sources, save notes and ask questions without switching apps.
AI Developments — Monday, November 16, 2026
Google Gemini
This Week in Google Gemini
Google Gemini is the AI that runs through Google’s apps, its Pixel phones and the tools coders use to build their own AI projects. This week Gemini had two big jobs at once. It moved deeper into your phone, and it shipped two new Gemini 4 models for the people who build apps.
The phone news came first. The November Pixel Drop, which started rolling out on November 10, brings “Gemini Intelligence” to the Pixel 10 family (but not the cheaper Pixel 10a). The fun part is Rambler, a voice-typing tool that cuts out every “um” and “like” and hands you a clean message to check before you send it. Gemini can now also run errands in Japan, the UK and India, like ordering groceries, booking a ride or grabbing a restaurant table. You can even describe a home-screen widget, say a countdown to your next trip, and Gemini builds it for you. On the Pixel 11 in the US, Call Screen got smarter too. Gemini now chats with unknown callers to find out what they want, a bit like a polite receptionist who screens your calls.
Under the hood, the developer side was even busier. Google’s Gemini API changelog shows Gemini 4 Carbon going generally available on November 10, at an introductory price of $2 per million input tokens (a token is roughly a chunk of a word) through November 30. Two days later came Gemini 4 Hydrogen, a cheaper model that even works on the free tier and adds an extra-hard “xhigh” thinking setting for tough problems. In between, on November 11, Google launched Gemini Decision 1, a model that does one thing: it makes a call. Show it text, a photo, audio, a video or a PDF and ask a yes-or-no or pick-a-category question, and it answers without writing a single word of chatter. Think of it as a referee instead of a commentator.
So what does it all add up to? Gemini is turning from a chatbot you visit into a helper that quietly does stuff, on your phone and inside other people’s apps. The Pixel Drop keeps rolling out over the next few weeks, so if your phone hasn’t picked up the new tricks yet, it probably will soon.
Meta AI — Llama models
This Week in Meta AI — Llama models
Quiet week for Meta AI — no major announcements in the past seven days, and no new Llama news either. That is mostly because Meta has moved its flagship AI from Llama to a new family called Muse. The last notable moves came just before our window opened. On November 5, Meta added real-time voice chats and expressive avatars to its Muse app, starting in the US. On November 6, it began rolling out new specialized models trained with Muse Wonder to both Muse and Meta AI. They are built to remember what matters from earlier chats and to keep going when a task hits a snag. Meta says Muse is now 2.4 times less likely to forget key details, though that number comes from Meta’s own tests.
Cursor
This Week in Cursor
Cursor is a code editor with AI agents built in. You describe what you want, and the agents write, test and fix the code with you. This week brought one fun new feature and one serious bit of homework.
The fun part landed on November 10, when Cursor Projects went multiplayer. Projects, which launched in September, used to be a solo space. Now you can invite teammates. Everyone gets a shared task board showing who is working on what and what each person’s agents are up to. Each person still has a private chat, but decisions and design docs live in a shared layer every agent can read. The neat twist is that the agents talk to each other. If yours and a teammate’s are about to change the same database, they can agree on an order first, like two movers deciding who carries the couch through the door. When someone new takes over the on-call shift, their agent can pick up the work already in progress. It is in beta and rolling out to everyone.
The serious part came on November 13, when Cursor posted an update on a recent security incident. An attacker got hold of an access key Cursor had given to Fireworks, one of its AI partners, and copied stored coding-session records, mostly from 2025. Cursor says users in Privacy Mode were not affected. Fewer than 0.05% of other users had their records exposed, and those people were contacted directly. The attacker also slipped fake copies of Anthropic’s Claude Code extension into Cursor’s extension store for about 11 hours on October 29. Cursor says the last attacker activity it saw was on November 1. Since then it has swapped out its passwords and keys, rebuilt the extension store and added scanning that catches leaked passwords.
The takeaway is simple. If Cursor emailed you, change the keys it listed. If you installed Claude Code versions 2.1.312 or 2.1.313 from Cursor’s store, follow Cursor’s October 29 cleanup steps. Then go enjoy coding with your friends, and your friends’ robots.
AI Developments — Sunday, November 15, 2026
Microsoft Copilot
This Week in Microsoft Copilot
Microsoft Copilot is the AI helper built into Word, Excel, PowerPoint and the rest of the Office apps that schools and offices run on. Last week we said the big September redesign was moving in. This week the boxes got unpacked, and the new stuff is a pile of small, practical “skills” that do the boring parts of homework and office work for you.
The handiest one lives in Excel. A new “clean data” skill, covered by Geeky Gadgets on November 10, hunts through a messy spreadsheet for duplicates, mixed-up capital letters and dates typed three different ways. Then it suggests fixes, and nothing changes until you say yes. Think of it as a tidy friend who asks before they rearrange your room. PowerPoint gets a “clarify my thinking” skill that quizzes you about what your talk is really for before it drafts slides. Word can now line up a jumble of pictures into a neat grid. Microsoft’s own release notes add more: Copilot can now borrow the look of a deck you attach, so new slides match your style. Excel answers now link straight to the cells Copilot changed. And Cowork, the mode where Copilot actually does tasks for you, can now click through websites in a separate Edge tab while your passwords stay on your own computer.
There is a catch for grown-ups paying the bills. Microsoft is rolling out usage-based billing in Copilot Chat and the Office apps this month, which works like a phone plan with a meter, per a November roundup from A Guide to Cloud. Autopilot, the always-on agent that keeps working after you close your laptop, still has no firm launch date. Microsoft’s notice, updated November 9, says the timing will come later.
That makes this coming week the one to watch. Microsoft Ignite opens Tuesday, November 17, in San Francisco and runs through the 20th. If Autopilot gets a real launch date, that is probably where we will hear it.
Canva AI / Magic Studio
This Week in Canva AI / Magic Studio
Quiet week for Canva AI / Magic Studio — no new AI features in the past seven days, and even a 14-day look turned up no new tools. The biggest news was about the company, not the product. On November 4, Canva announced a restructure affecting about 1,271 people, roughly 22% of its staff. CEO Melanie Perkins said growth had slowed for the first time in a decade, and the company pointed to new technology changing how people make things. The last big AI launch was on October 29: Canva Brush, a model built to edit designs you already have rather than draw new pictures. It rolled out first to Canva Pro subscribers, one week after Canva Craft, its image-making model.