China’s free GLM 5.2 model, Claude lawsuit, and 16 more big AI updates this week

This week in AI is wild: the US just blocked one of the most powerful models for most of the world, a Chinese company responded by dropping a free rival you can run on your own laptop, Anthropic is facing a lawsuit over Claude usage limits, and Midjourney quietly stepped into healthcare with a full-body scanner. On top of that, OpenAI, NVIDIA, Perplexity, and others shipped features that change how you build, automate, and design with AI.

GLM 5.2: China’s open model that rivals Claude and GPT

A Chinese company called Z.ai has released GLM 5.2, a large language model that’s open, cheap, and surprisingly close in quality to the very best proprietary models.

Most frontier models from OpenAI, Anthropic, and Google are closed: you access them through an API or app, they can change prices or rules at any time, and they can cut you off completely. We’ve already seen this with Anthropic’s most powerful model, Claude Fable 5, which was recently banned for users outside the US.

GLM 5.2 is different in four important ways:

First, it’s open weight. You can download the full model and run it on your own machine. No subscription, no platform lock-in, and nobody can ban you from using it locally.

Second, it’s released under an MIT license. That means you can use it for personal projects or inside a business, including commercial products, without paying licensing fees.

Third, it supports a 1 million token context window. In simple terms, you can feed it huge projects—entire codebases, long documents, or multi-file apps—and it can keep track of everything without forgetting how things started.

Fourth, it offers two thinking modes, High and Max, with Max going all-out for harder, more complex tasks.

In blind evaluations, GLM 5.2 ranked as the second-best AI model in the world, just behind Claude Opus 4.8, while costing about six times less. On very long, complex tasks, top paid models like Claude still pull ahead, but GLM 5.2 is close enough that the price difference becomes a big deal.

If you want a deeper, hands-on walkthrough, there’s also a dedicated guide on how to use GLM 5.2 inside Claude Code and where it actually beats Opus.

Anthropic lawsuit: are AI usage limits misleading?

Anthropic, the company behind Claude, is facing a lawsuit that could change how AI companies disclose usage limits.

A customer named Carl Kahn signed up for Claude Max 5× and Max 20× plans, expecting five or twenty times the usage of the regular Pro plan. According to his complaint, the real limits were far lower than advertised. In one example, a single five-hour session allegedly consumed 15% of his entire weekly limit—meaning a $200 plan could effectively last only a few days under heavy use.

He filed the case in California and is seeking class action status, which would allow anyone who bought a Max plan since launch to join. Anthropic has declined to comment so far.

Almost every AI provider today uses fuzzy, token-based or session-based caps that most users don’t fully understand. If this case moves forward, it could force AI companies to be much more explicit about what you’re actually buying before you pay.

Midjourney Medical: from AI art to body scanning

Midjourney, best known for its stunning AI-generated images, is stepping into healthcare with a new initiative called Midjourney Medical and a hardware prototype called the Midjourney Scanner.

The concept is very different from traditional hospital imaging. You step onto a platform, get lowered into a shallow pool of warm water, and a ring of thousands of tiny sensors scans your body using sound waves. The goal is to avoid the claustrophobic experience of MRI machines and make the whole thing feel more like a spa than a hospital visit.

The company plans to open a space in San Francisco that combines saunas, cold plunges, and scanning rooms. The long-term vision is a 60-second, low-cost scan that could be deployed in 50,000 locations worldwide.

Reality check: it’s still early. The scanner is a prototype, each scan currently takes about 20 minutes, and it still needs FDA approval. But it’s a bold signal that AI-native companies are starting to invest their profits into ambitious health and hardware projects, not just software.

AI talent wars: a founding father of transformers joins OpenAI

Noam Shazeer, one of the key researchers behind the transformer architecture, has left Google for OpenAI.

Back in 2017, Shazeer co-authored the landmark paper “Attention Is All You Need”, which introduced transformers—the architecture that powers almost every major AI model today, including ChatGPT, Claude, and Gemini. That’s why he’s often called one of the “founding fathers” of modern AI.

He worked at Google for years, left to start his own company, and then in 2024 Google reportedly spent around $2.7 billion to bring him back and put him in charge of Gemini, their flagship AI.

Less than two years later, he’s moved to OpenAI as lead of architecture research, where he’ll help design how future OpenAI models are built from the ground up. It’s a clear sign of how intense the AI talent war has become: a single top researcher can be worth more than many entire startups.

ChatGPT gets instant camera input and scheduled tasks

OpenAI quietly rolled out two quality-of-life upgrades for ChatGPT that make it more useful in everyday life.

Instant camera answers on mobile

On iOS, you can now open ChatGPT, tap the camera, and point it at something—like a stain on your shirt—and ask a question. The photo lands in the chat instantly and the model responds without the lag users were used to. The camera is now integrated directly into the conversation flow.

This is currently iOS-only, with Android support expected later.

Scheduled tasks that run in the background

ChatGPT can now run tasks on a schedule without you having to remember to ask. For example, you can say: “The day before every Portugal match, send me a scouting report on the other team.”

Once set, ChatGPT prepares the report in the background and sends you a notification at the right time. There’s also a new schedule page where you can see, edit, or cancel all your recurring tasks in one place. It’s a step toward AI that behaves more like a proactive assistant than a passive chatbot.

Physics Wallah’s bilingual AI tutor for Indian students

In India, one of the biggest AI education stories is coming from Physics Wallah, a platform used by over 36 million students preparing for exams like JEE and NEET.

Their AI doubt solver just gained a voice interface that speaks in natural “Hinglish”—a mix of Hindi and English that mirrors how many students actually talk and think.

Here’s how it works: a student scribbles a question on a digital blackboard, even with messy handwriting. The AI reads the handwriting, understands the question, and then explains the solution out loud, step by step, like a patient teacher sitting next to you at 2 a.m.

The voice is powered by ElevenLabs, a text-to-speech tool known for its human-like audio. After launching voice, Physics Wallah saw students asking 3× more questions per session and staying engaged 2.4× longer within just two weeks.

NVIDIA’s AI animator for games and robots

NVIDIA has introduced an AI system that can generate character movement in real time, potentially replacing huge chunks of traditional animation pipelines.

Today, animators often create character motions by hand—walking, jumping, vaulting, sitting—frame by frame or through complex motion capture setups. It’s time-consuming and expensive.

NVIDIA’s new AI, built by a team called Motion Metrics, can generate these movements on the fly. In a demo, a character navigates a 3D space, picks up a sword, vaults over a bench, sits down, and switches between different movement styles like zombie, injured, or skipping—all generated live by the model.

This isn’t just for games. The same underlying system is also being used to control humanoid robots, meaning the AI that animates a virtual character can also drive a real robot’s body in the physical world. Researchers are already using it to move real robots in labs.

OpenRouter Fusion: combining multiple models to rival Fable 5

OpenRouter, a platform that lets you access many AI models through a single interface, has launched a feature called Fusion. The idea is simple but powerful: instead of picking one model, you can run multiple models on the same task and combine their strengths.

Where one model is weak, another might be strong, so the fused output can be better than any single model. In tests across 93 different tasks, a panel of four cheap “budget” models fused together nearly matched the performance of Claude Fable 5—at about half the price.

For users locked out of Fable 5 due to regional bans, Fusion offers a way to get similar intelligence levels by blending multiple accessible models.

Perplexity Brain: an AI that remembers your work

Perplexity has introduced Perplexity Brain, a feature that lets the AI remember your ongoing work instead of starting from zero every time.

Normally, each new task you give an AI is isolated. You have to keep re-explaining your project, your documents, your context, and your preferences. Brain changes that by quietly building a map of what you’re working on—documents, previous answers, and relevant links.

For example, if you’ve already launched a product and loaded your docs and channels into Brain, the next time you ask about a launch plan, it already knows your previous strategy, assets, and constraints. Every entry is readable, searchable, and traceable.

Perplexity says that with Brain, the second time you ask for something, the AI gets it right 25% more often and runs 13% cheaper, because it can reuse context instead of recomputing everything from scratch.

GenSpark Agent Base: one AI tool instead of a dozen SaaS apps

GenSpark, known for AI that can build slide decks and even make phone calls, has launched Agent Base, a platform designed to replace many of the separate tools teams use today.

Most teams juggle a mess of apps: one for CRM, another for marketing, spreadsheets for tracking, separate dashboards for analytics, and so on. None of them talk to each other perfectly, and all charge monthly fees.

With Agent Base, you describe what you want in a single sentence—like “Build a dashboard for this month’s sales.” The AI pulls in your data, analyzes it, and creates a working dashboard with charts, revenue numbers, win rates, and more. You can then talk to the system to update it on the fly.

Instead of stitching together many SaaS tools, Agent Base aims to be a single, adaptive system that molds itself around how your team actually works.

Lovable’s point-and-draw editing for apps and sites

Lovable is a tool that lets you build full apps or websites from text prompts. Its latest update makes editing much more visual and intuitive.

Previously, if you wanted to change something, you had to describe the edit in words. Now, Lovable adds a toolbar with several modes, including a standout feature called Annotate.

In Annotate mode, you can draw directly on your live website: cross out a button you don’t want, circle an element to change, or draw an arrow to move a section. The AI reads your marks and applies those edits automatically.

You can also:

• Edit text directly on the page and have it saved automatically.
• Select specific text boxes or visual elements to tweak.
• Add comments, tag teammates, and send instructions straight to Lovable for collaborative editing.

It’s a step toward building and editing software by pointing and sketching, not just typing prompts.

Claude Design + Replit: from prompt to live app

Anthropic has upgraded Claude Design to make it more practical for real-world work and tightly integrated it with Claude Code and Replit.

First, Claude Design can now learn your existing style. Point it at your website or design files, and it automatically picks up your colors, fonts, and visual language.

Second, you can edit directly on the canvas. Click elements to move, resize, or align them visually instead of typing out every change as a prompt.

Third, Claude Design and Claude Code now work together, so you can switch between design and implementation without starting over. When you’re done, you can export designs to PDF or PowerPoint, or send them straight into tools you already use.

One of the most powerful new flows is the integration with Replit. For example, you can tell Claude: “Design a sleek skincare shopping app called Luna Skin Care with a soft lunar aesthetic and a calm minimalist feel.” Claude designs the full interface. Then you hit share, choose Send to Replit, and Replit turns that design into a real, runnable app you can build on and deploy.

The line between describing, designing, and shipping an app is getting much thinner.

Claude Code live artifacts: dashboards that update as the AI works

Claude Code can now turn its work into live, shareable pages—called artifacts—that update in real time as the AI keeps coding.

Imagine you ask Claude Code to investigate why users are dropping off after a recent product update. It pulls the data, builds a dashboard, highlights the problem, and suggests fixes. That dashboard is a live page you can share with your team via a link.

As Claude continues to refine the code or analysis, the page updates itself automatically. Teammates can open it on their phones or laptops and always see the latest version without you having to re-explain anything.

It effectively turns every coding or analysis session into a live, collaborative artifact that the whole team can follow.

OpenAI Codex record-and-replay: teach once, reuse forever

OpenAI has added a record and replay feature to Codex that makes automating repetitive tasks much easier.

Instead of writing a detailed prompt describing every step, you simply perform the task once while Codex watches. For example, you could record yourself uploading a video to YouTube Studio: adding the title, description, uploading the file, choosing settings, and handling pop-ups.

Codex observes the entire flow and turns it into a reusable skill. Next time, you just tell Codex to “upload the video,” and it replays the same steps automatically. You can open the skill, see what it learned, and edit it if needed.

It’s like teaching your computer by demonstration instead of instructions.

GLM 5.2 vs Claude and Opus: real-world tests

Back to GLM 5.2: how does it actually perform against Claude Fable 5 and Claude Opus 4.8 in real tasks? Three tests—roadmaps, interactive sites, and games—show how close it really gets.

Test 1: AI learning roadmap website

The first test asked each model to build an AI learning roadmap as a website.

GLM 5.2 didn’t just start coding. It first asked three clarifying questions: your background (e.g., student), how much time you have per day (say, 30 minutes), and what you want to use AI for (e.g., work). Then it generated a full skill tree and a 90-day, week-by-week plan with dropdowns you can adjust. It also included real learning resources with clickable links to courses.

Change one input—for example, from 30 minutes a day to 1 hour—and the entire plan recalculates itself around the new time budget.

Claude Fable 5 produced a similarly detailed week-by-week plan from week 1 to week 13, starting with foundational material like Andrej Karpathy’s “Intro to LLMs,” then moving into AI engineering and ending with a build phase where you ship something real. It added a progress tracker and a skill tree that shows what unlocks next. Clean and very usable.

Claude Opus 4.8, using the same prompt family, generated a nice title screen with a “Build My Roadmap” button—but clicking it did nothing. The full week-by-week plan simply wasn’t wired up. Same model family, slightly older version, but a very different (and less useful) result.

Conclusion: GLM 5.2 can go toe-to-toe with Fable 5 on this kind of structured, interactive planning task, while Opus 4.8 lagged behind in execution.

Test 2: interactive “One Wish Below” website

The second test asked each model to build an interactive website themed around “One Wish Below” from Obsession.

GLM 5.2 returned a fully interactive page: as you scroll, animations play; you type your wish into a box, click to seal it, and the page reacts. It added a call-to-action button, custom fonts, drifting particles, and a sound toggle—lots of small interactive touches that make it feel alive.

Claude Opus 4.8 produced a clean layout with good typography and a working wish box. When you pull a rope, paper glitter falls across the screen—a nice bit of flair.

Claude Fable 5 delivered the most beautiful design and writing, with a strong mood and aesthetic. But it lacked interactivity, so it felt more like a finished poster than a page you can play with.

All three did well, but GLM 5.2 offered the richest interactivity—the most elements you can actually click, trigger, and experience.

Test 3: building a complete 2D action game

The third test pushed all three models harder: “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 sound effects. Think old-school side-scrolling shooter.”

GLM 5.2 delivered a playable game where a pixel-art commando runs, jumps between platforms, and fires nonstop. Enemies spawn continuously, each shot has a sound effect, and your score and three lives are displayed at the top, just as requested.

Claude Fable 5 built a game called “Jungle Strike,” also fully playable. The soldier runs, jumps, and shoots, enemies attack, and there’s a rapid-fire power-up that makes your gun shoot much faster. The score climbs as you play—solid execution.

Claude Opus 4.8 created another smooth-running game with a standout twist: a spread-shot power-up that turns your weapon into a multi-bullet spray, making it easier to clear enemies and adding variety to the gameplay.

All three passed this test impressively. GLM 5.2 clearly belongs in the same league as the top Claude models for game-like coding tasks.

Pricing: GLM 5.2 vs Claude Opus

Performance is only half the story. The other half is cost.

Claude Opus 4.8 pricing is roughly:

• About $5 to read 1 million tokens.
• About $25 to write 1 million tokens.

GLM 5.2, by comparison, is around:

• $1.20 to read 1 million tokens.
• $4.10 to write 1 million tokens.

That’s roughly five times cheaper for similar quality work, and often with faster responses. For teams running large workloads, that difference can translate into massive savings.

How to start using GLM 5.2 (and what to watch out for)

There are a few ways to get started with GLM 5.2, depending on how much control and privacy you need.

1. Use it instantly in your browser

The easiest option is to go to chat.z.ai. You can start using GLM 5.2 right in your browser with zero setup. You’ll get a set of free tries and can immediately start building websites, tools, or experiments.

2. Use it via OpenRouter in coding tools

If you want to integrate GLM 5.2 into a real coding environment without spending much, you can use it through OpenRouter. Create a free OpenRouter account, generate an API key, and drop it into a tool like OpenCode. From there, you can select GLM 5.2 from a list and use it like any other model.

3. Data privacy: China-based servers vs local runs

One important caveat: when you use GLM 5.2 through the cloud (chat.z.ai or any hosted API), your prompts and data are processed on Z.ai’s servers, which are located in China. That means any private documents, client data, or proprietary code you paste in will pass through machines you don’t control, under a different regulatory environment.

If that bothers you, there’s a solution: because GLM 5.2 is open weight, you can download the model and run it locally on your own hardware. In that setup, all computation happens on your machine and nothing leaves your computer.

Tools like Ollama make this easier by handling the heavy lifting of downloading and running large models. You’ll need a powerful machine—GLM 5.2 is big—but in return you get full control and maximum privacy.

Also note: some people mention that NVIDIA offers this model for free. That’s only half true. NVIDIA’s free option currently runs the older GLM 5.1, not GLM 5.2. So don’t expect to get the newest model that way just yet.

Why this week matters for AI builders

Zooming out, the pattern is clear: a year ago, building polished websites, interactive experiences, or full games with AI usually meant paying for the most expensive proprietary models. Now, a free, open model like GLM 5.2 can match or closely approach that level for a fraction of the cost.

At the same time, we’re seeing:

• Legal pressure on AI companies to be more transparent about pricing and limits.
• AI-native companies like Midjourney reinvesting profits into ambitious hardware and health projects.
• Tools like OpenRouter Fusion, Perplexity Brain, and Codex record-and-replay making AI more collaborative, persistent, and automated.
• Design-to-code flows (Claude + Replit, Lovable) that compress the entire build pipeline into a single conversation.

If cost or complexity was the thing stopping you from building a website, app, or internal tool with AI, that barrier is rapidly disappearing. Open models from China and elsewhere are becoming serious contenders, and you can already see that trend in other areas too—like China’s wave of free AI video generators that don’t even require sign-up.

The next phase of AI won’t just be about who has the single smartest model. It will be about who can combine open and closed models, automate workflows, respect user data, and give builders the most freedom at the lowest cost.