Transcription
Last week, the US banned the most powerful AI in the world for everyone who's not American. Now, a Chinese company just dropped an AI that can compete with it, and it can be used completely free. Anyone can download it, run it on their own laptop, and nobody can restrict you from using it.
But that's not even the wildest story this week. Anthropic just got sued because someone figured out that Claude's limits were nowhere close to what was advertised. They were running out way faster than what he paid for, and you need to watch to find out what this means for you.
And then there's MidJourney, the company that makes those stunning AI images. Well, they just switched lanes completely and built their first ever body scanner. Wait till you see why. It'll blow your mind.
There are 16 such updates that you should know about this week. And right at the very end, we run the free model we told you about in the start, GLM 5.2, head-to-head against the banned Claude Fable 5. And honestly, the result will surprise you.
Now, if any of these updates get a little technical, don't worry. The timestamps are in the description. Jump to whichever you want to watch, and the full step-by-step guide, plus a method on how to use GLM 5.2 for free, is waiting for you in our free WhatsApp community called Staying Ahead.
Now, let's dive into the video. A Chinese company called Z.AI AI just released a model that almost beats frontier models, and it's free for everyone. Let me explain why that matters. Most of the top AI models, the ones from OpenAI, Anthropic, Google, are locked behind their companies. You can use them, but you don't own them. They can change the price, add rules, or even cut off access. We literally just saw that happen with Claude's Fable 5 getting banned for people inside and outside the US. GLM 5.2 is public. Anyone can download it and run it on their own computer. No subscription. Nobody can shut it off or ban it. And this isn't some weak alternative just because it's free. In a blind test, GLM 5.2 ranked the second best AI in the world, beaten only by Claude Opus 4.8, and it costs six times less. On the hardest, longest tasks, the top paid models like Claude still pull ahead. So, it's not the ultimate winner, but it's pretty close. We were so fascinated by it, so we decided to test it out and make a tutorial around it. But you have to wait till the end to catch that because we've got some more important AI updates to catch first.
Anthropic just got hit with a lawsuit. A customer named Carl Khan signed up for Max five times and Max 20 times plans, expecting five or 20 times the usage of the regular pro plan. But according to his lawsuit, the real limits were nowhere close to what was advertised. One single 5-hour session used up 15% of his entire weekly limit, which is hardly 3 days for a $200 plan. He filed the case in California and wants it certified as a class action. So, anyone who bought a max plan since it launched could join in. Anthropic, however, declined to comment on this matter. Every AI company runs on these fuzzy usage caps that almost nobody fully understands. So this case could force them to tell you exactly how much you're getting before you pay.
Midjourney, the company you know for making AI images, just announced its debut into healthcare called Midjourney Medical. They are building the Midjourney Scanner, actual hardware. And here's the idea. You step onto a platform, it lowers you into a shallow pool of warm water, and a ring of thousands of tiny sensors scans your entire body using sound waves. No need to use claustrophobic machines like in an MRI. But why is an AI image company building a body scanner? The founder says making AI images earns them enough to fund crazy bets like this. And his real goal is improving human health, not just making pictures. And this whole experience of theirs feels like a spa, not a hospital. They're opening a place in San Francisco with saunas, cold plunges, and scanning rooms. The dream is to do a 60-second scan that costs a few dollars with 50,000 of these worldwide. Right now, this is still a prototype. It still needs FDA approval, and a single scan currently takes about 20 minutes, not 60 seconds. But this is a bold vision and a step forward. Let's see whether it becomes a reality.
The founding father of AI just left Google for OpenAI. His name is Noam Shazir, and Google paid $2.7 billion to keep him in their company. Back in 2017, Shazir was one of the people who wrote a research paper called "Attention Is All You Need." That paper invented the core idea that basically every major AI today is built on. ChatGPT, Gemini, Claude, all of them. So when people call him one of the founding fathers of modern AI, they're not exaggerating. He used to work at Google, left to start his own company, and in 2024, Google paid around $2.7 billion to bring him back and put him in charge of Gemini, their top AI. Now, less than two years later, he's walking out the door and into OpenAI, their biggest rival. OpenAI made him lead of architecture research, basically putting him in charge of designing how their future AI models are built from the ground up. So the person Google spent billions to win back just joined the other team. That's how fierce the fight for AI talent has gotten. The smartest people in the field are worth more than entire companies, and everyone's trying to poach them from everyone else.
Earlier, uploading a picture in the ChatGPT app was a hassle. You had to wait until it loaded. Now watch this. You're out eating tacos, you spill something, and there's a stain on your shirt. You open ChatGPT, tap the camera, point it at the stain, and ask, "How do I get rid of this stain?" And it answers right away without any lag. The camera opens right in the chat, the photo lands instantly, and you get your answer. One catch for now, it's iOS only. Android users need to wait for their turn.
While everyone's glued to what's launching in Silicon Valley, the biggest AI education story this week is sitting right here in India. Physics Walla, the platform 36 million students use to crack JEE and NEET, just gave its AI doubts a voice. And it speaks in the language of most Indian students, a combination of Hindi and English, the exact way you'd explain a physics problem to a friend. Here's how. A student scribbles a problem on a digital blackboard, say calculating work done in a question, messy handwriting and all. The AI reads the handwriting, understands the question, and explains the whole solution out loud, step by step, like a patient teacher sitting next to you at 2:00 AM. That voice runs on 11 Labs, the text-to-speech tool that makes AI sound more human. Once voice went live, the students who used it asked three times more questions per session and stuck around 2.4 times longer.
Nvidia just built a single AI that replaces decades of animation technology. Right now, every single movement a game character makes has to be built by hand, one at a time, by an animator. Picture someone drawing every tiny motion frame by frame. It takes weeks. Motion Bricks developed an AI that creates these movements on its own in real time. Watch this demo. A character navigates a space, picks up a sword, vaults over a bench, sits down, and switches between zombie, injured, and skipping styles. Every single frame generated live by the AI. And this isn't just for games. This same brain runs Nvidia's humanoid robots. So, the same AI behind a game character's movements is what runs a real robot's body. Insane, right? Researchers are already using it to move real robots.
OpenRouter, which is basically a single doorway to every AI model, just dropped something called Fusion. And the timing is perfect. Ever since Claude's most powerful model, Fable 5, got banned for everyone outside America, people have been hunting for that same level of intelligence. Fusion is finally it. Normally, you'd pick one AI model and you are stuck with whatever output it provides. Fusion, just as the name suggests, fuses multiple AI models and runs the same task at once. This fusion hits Fable 5 level performance because where one model is weak, another covers for it. So, the combined answer ends up stronger than any single model alone. They ran it across 93 different tasks. And here's what's insane. A panel of four cheap budget models fused together nearly matched Fable 5 at half the price. When experimented with single AI models, the results were satisfactory.
ChatGPT just introduced a better way to schedule tasks. ChatGPT can now work in the background without waiting for you. All you need to do is just schedule what you want beforehand. If you type, "The day before every Portugal match, send me a scouting report on the other team." Done. Then the day before the game, a notification just shows up with the full report, key players, weaknesses, all of it. Set it once and it runs on its own. There's even a new scheduled page to manage everything in one place.
Most AI remembers who you are. Perplexity now remembers what your work is. It's called the Perplexity Brain. Normally, every time you give an AI agent a task, it starts from zero. So, you keep repeating yourself. Perplexity Brain quietly builds a map of everything you've worked on. And every new job starts with all of that already loaded in. Say you've launched a product before, the next time you ask, it already knows your docs, your channels, and how you did it last time. And every entry is readable, searchable, and traceable. So the second time you ask for something, Brain gets it right 25% more often, remembers more of what matters, and runs 13% cheaper.
Genspark, the company whose AI builds full slide decks and even makes phone calls for you, just launched a tool that replaces all the apps you pay for every month. It's called Agentbase. Think about how most teams work today. One app for customers, another for marketing, a Google Sheet for tracking, and an Excel file for the numbers, a dozen separate tools, none of them talking to each other, all charging you every month. And none of them quite fitting how you actually work. With Agentbase, you just type one sentence, and it creates a full working tool right there. If you asked it to build a dashboard for this month's sales, it pulls the data, thinks through it, and instantly puts up charts, revenue numbers, win rates, the whole thing. Then you just talk to it, and it updates on the fly. It pulls in your real data and turns all of that scattered stuff into one clean system that's built around you.
What if you could point at something for the change to be made instead of typing it out? That's exactly what Lovable just did. Lovable lets you build a full app or website just by typing prompts. And until now, editing worked the same way. With the new update, you get a toolbar that lets you edit by pointing and drawing instead. It has four modes, but the best one is annotate. You scribble right on your live website, like cross out a button you don't want, circle something to change, draw an arrow to move a section, then the AI reads your marks and makes those edits for you. You can fix text right on the page, and the change saves itself automatically, or select any specific text boxes or visual elements to edit. Now you can also add comments, tag teammates for collaboration, or send instructions directly to Lovable.
Anthropic just made Claude Design a lot more useful for actual work. First, you can point Claude at your existing website or design files, and it picks up your colors, fonts, and style on its own. Second, you can also edit directly on the canvas. Click anything to move it, resize it, or line it up right on the screen instead of typing out every change. Also, Claude Design and Claude Code now work together. So, you can design, build, and go back and forth without starting over. So, when you're done, export straight to PDF or PowerPoint or send it to tools you already use.
You can now design something in Claude and send it straight to Replit, the tool that builds and runs a full app just from your description, to turn it into a real working app. If you type into Claude, "Design a sleek skincare shopping app called Luna Skincare with a soft Luna aesthetic and a calm minimalist feel." Claude designs the whole thing. Then hit share, pick send to Replit, and it's gone over to Replit where it gets turned into an actual app you can build on and put online. So, the design and the building aren't two separate jobs anymore. You describe it, design it, and ship it all in one flow.
Claude Code can now turn its work into a live page so your whole team can watch updates in real time. When an AI does a bunch of work, you usually have to stop and explain it all to your team. Not anymore. If you ask Claude Code to figure out why users are dropping off after a recent update in seconds, Claude pulls the data and builds a live dashboard and spots the problem. Then it suggests fixes. You can share the link to your team, and they can open it right on their phone. As Claude keeps working and changing the code, that same page updates itself live. Anyone can see the latest version. It's like turning every coding session into a dashboard your whole team can follow. And if you want to see what else these Claude artifacts can do, we have made a full video breaking them down. Go check it out.
OpenAI just added something to Codex that could make prompting your AI easier. Normally, to get an AI to do something, you have to describe every step. Now, you just show it once. That's the whole idea behind Codex's new feature, Record and Replay. Say you record yourself uploading a video to YouTube Studio. You hit record, add the title, description, upload the file, pick the settings, and handle the pop-ups. Codex quietly watches the whole thing sitting on your computer. When you're done, it turns everything it just saw into a reusable skill. Next time, just tell Codex to upload the video, and it runs the exact same steps on its own. You can open up that skill, see what it learned, and edit it. So, you teach it once, and it handles it from then on.
Okay, now let's get to the most exciting part. The brand new AI model out of China that's on par with GPT 5.5 and Claude Opus 4.8, but is only a fraction of the cost. It's called GLM 5.2. It's built by a company called Z.AI, and we're going to test it right now. We'll take the exact same prompts we gave Claude Fable 5 and Opus 4.8, word for word, and watch it build full websites, a game, and a design rebuild on its own. Before we begin, the step-by-step guide on every free way to run GLM 5.2, plus a bonus on running it for free, is waiting in my free WhatsApp community. Grab it using the link in the description to follow along. And now, let's dive into the tutorial.
Before we watch it work, let's give you some quick context about this. First, it is open weight. In plain English, that means anyone can download the whole model and run it on their own machine. Second, it has an MIT license, which means it's free to use even inside a business. Third, it holds a 1 million token memory, which in plain English means we can hand it an entire project and it will not forget how things started. And fourth, it comes with two thinking modes: high and max, where max is the one that goes all out.
Now, let's test it out. We are going to use the same prompts for all three models: Opus 4.8, Claude Fable 5, and GLM 5.2. So the comparison is fair. Let's start with GLM 5.2 and give it the first prompt, which is basically asking AI to build us an AI learning roadmap as a website. Watch what it does before it builds anything. It stops and asks us three questions. First, what's your background? Let's pick student. How much time do we have? Let's be honest and say 30 minutes. And what do we want to use AI for? Let's go with work. And look at what comes back. It builds a full skill tree and a 90-day plan laid out week by week with drop-downs that you can change as per your needs. You also get proper resources. If you click those, you get redirected to the courses. Now, here's the part to watch closely. Change one input, 30 minutes a day to 1 hour a day, and the whole plan rebuilds itself around the new number.
Now finally, try running Fable 5 with the same prompt like GLM 5.2. It also builds a plan week by week, all the way from week 1 to week 13. It starts with "Intro to LLM talk by Andre Karpathy." Then it moves into AI engineering and ends with a phase where learning stops and something actually gets put out into the world. There's a progress tracker down the side so you can tick things off as you go. And there's a skill tree that shows what unlocks next. The whole thing is clean and it just works.
Let's try it next on Claude Opus 4.8. It runs, gives us this file, and here is the result. It's a nice title screen with a button called "Build My Roadmap." We click the button, and literally nothing happens. Nothing opens. The full week-by-week plan we just showed you isn't there. We're using the same family of AI, only a few months older, but are getting a completely different result. So, in summary, GLM 5.2 can run neck-and-neck with Fable 5, and Opus 4.8 lags behind.
Okay, now let's push GLM 5.2 a little harder. Let's give it a prompt to make an interactive website for "One Wish Willow" from Obsession. Here's what GLM came back with. As you scroll, the page animates. You whisper a wish by typing it in, click to seal it, and the whole thing reacts. It even added a CTA button on its own, plus custom fonts, little particles drifting around, and a sound toggle. It's pretty cool.
For comparison, here's Opus 4.8 on the same prompt, and it's good. It's got a clean layout, nice typography, and the wish box functions like it should. When you pull the rope, paper glitter starts dropping on the screen. Then, Fable 5. This one is beautiful. The writing is lovely. The mood and the design are spot-on. The only thing missing is the interactivity, so it feels more like a finished poster than something you can click around in. While all three models have great results, GLM 5.2 gives us the most to actually play with, more moving, clickable, reacting parts than either Claude model.
Finally, for the third test, we wanted to find out what happens when you ask an AI to create something far more complex, like an actual playable video game. Let's start with GLM 5.2 again and give it this prompt: "Build a complete 2D action game in one file with a commando-style soldier, a jungle level, power-ups, three lives, a score counter, and even sound effects. Think old-school side-scrolling shoot-em-up action." This is what we got back. A little pixel art commando runs, jumps between platforms, and fires non-stop. Enemies keep coming. Every shot has its own sound effect. And your score and three lives sit right at the top, exactly as the prompt asked.
Let's now put the same prompt on Fable 5. It built us a game called "Jungle Strike," and it's fully playable. Let's try it. The soldier runs, jumps between platforms, and shoots, and enemies keep coming at him. We grab a rapid-fire power-up, and the gun changes. It starts firing way faster. The score climbs as we play. It's a solid game.
Finally, let's move on to the older Claude, Opus 4.8. When given the prompt, here's the game it built. The game runs smoothly, but the coolest part was a spread-shot power-up. Instead of shooting a single bullet, your weapon suddenly sprays multiple shots across the screen, which makes clearing enemies much easier and gives the game a bit more variety as you play. All three are fantastic here, and GLM 5.2 went toe-to-toe with Claude's models, so it can build well.
The next question any sensible person asks is, what does this cost? Well, Opus 4.8 costs about $5 to read a million tokens, basically a million chunks of words, and $25 to write a million back. GLM 5.2 does the same work for $1.2 and $4.1 respectively, which is around five times cheaper. The same quality of work for a fraction of the bill, and often times faster results, too.
So, how do you get started with this tool? You've got a few options. The easiest and free way is to head to chat.z.ai. It's right there in your browser with zero setup. You get a set of free tries and can start building instantly. Next, if you want to use it inside a real coding setup without spending much, you can use OpenRouter. All you need to do is make a free account, generate a key, and drop it inside a tool called OpenCode that's also free. Basically, it's like copy-pasting GLM 5.2 and tokens from a list.
There is one thing you need to know before using this tool on the cloud. Whenever you use an AI tool through a website or an app, what you type doesn't stay on your screen. It gets sent off to a computer somewhere, the model does its work there, and the answer comes back to you. For the cloud version of GLM 5.2, that somewhere is Z.AI's own servers, and those sit in China. So, if you paste in something private, a client document or code you'd rather not hand over, it has traveled through a machine you don't control in a country with its own rules about data. If that bothers you, here's the way out. Since this model is open weight, anyone can download it and run it themselves. So, instead of sending your work off to their cloud, you can run GLM 5.2 on your own laptop, and it does all its thinking right there on your machine. Nothing ever leaves your computer. To do this, you can download the whole model onto your own machine for free with a tool called OAMA. We've put together a clean step-by-step guide on exactly how to do this. It's free. It's sitting in our community, and the link is below. Go grab it. One honest heads up, this thing is massive, so you'll need a seriously powerful computer to run it.
One last thing I want you to note. A lot of people are saying Nvidia gives this model for free, and it's half true. Nvidia's free version is running the older GLM 5.1 right now, not GLM 5.2 yet. So, don't go chasing that expecting to run the new model.
Now, zoom out for a second. A year ago, building something this polished meant paying for the most expensive model on the planet. Today, a free, open model is matching it. So, if you've ever wanted to build a website, a small app, or a little tool for your work, and the cost was the thing stopping you, that wall has come down. If you like such free tools, I just made an entire video masterclass on all of Google's free tools that you should know about and can use immediately. Go watch that next.
Now, we upload such AI updates and tools tutorials every week as soon as they drop. To catch the next one and stay ahead of the curve, click on that subscribe button. I hope to see you in my next video.