March 25, 2026: OpenAI Kills Sora, Anthropic's Claude Takes Control of Your Mac

Today's key AI stories

• OpenAI just killed Sora. The AI video model is being shut down. The reason? OpenAI wants to build a super app instead.
• Anthropic's Claude can now control your Mac. It can click buttons, open apps, and type for you. This is the most ambitious AI agent yet.
• Google has a new efficiency trick. TurboQuant can compress AI models by 6x without losing accuracy.
• Experts are warning about AI agent risks. A new eBook asks: are we ready to give AI real autonomy?
• ChatGPT is now a shopping app. You can browse products, compare prices, and buy directly in chat.

OpenAI Breaks Up With Sora

OpenAI is shutting down Sora. The AI video model that wowed the world in 2024 is gone. The app is gone. The API is gone. No specific shutdown date was given. The company just posted on X: "We're focusing on a super app."

This is a big shift. Sora launched just 10 months ago as Sora Turbo. It could generate 60-second videos with realistic scenes. But competitors like Runway, Luma, and Chinese companies Kling and Minimax already caught up.

The timing is strange. OpenAI released updates to Sora all week. Then boom. Gone. Disney had pledged $1 billion to bring Disney characters to Sora. That deal is now canceled.

Why kill Sora? OpenAI wants to consolidate. Think super app with ChatGPT, Codex, and everything else in one place. This aligns with their new OpenAI Foundation. The Foundation plans to invest $1 billion in life sciences, jobs, and AI resilience.

The message is clear. OpenAI is moving away from fun AI content. They're moving toward serious business. Curing diseases. Economic impact. This is a different OpenAI.

What it means

The AI video race is over. OpenAI quit. They decided it's better to be the super app than the video tool.

This matters for creators. Sora was supposed to change video forever. Now that future is gone. Competitors will fill the gap. But the leader walked away.

Anthropic's Claude Now Controls Your Mac

This is the big one. Anthropic just gave Claude the ability to control your computer. Not metaphorically. Literally. It can click buttons. Open apps. Type in fields. Navigate your desktop while you grab coffee.

The feature is available now for paying subscribers. It works inside Claude Cowork and Claude Code. You can even control it from your iPhone using Claude Dispatch.

How it works. Claude checks for direct connectors first. Gmail, Slack, Google Drive. Those are fastest. If no connector exists, it uses the Chrome browser extension. Only as a last resort does it interact directly with your screen. Clicking. Typing. Scrolling.

This matters. It's a three-tier system. Connectors are fastest. Browser is slower. Screen interaction is slowest and most fragile. But it works with any app.

Early tests show mixed results. One review found it works about half the time. Success for simple tasks. Failure for complex multi-step workflows. One tester called it "a step in the right direction" but said "it's not good enough to rely on when you're away from your desk."

The security risk

Here's the scary part. Claude is interacting with your real desktop. Not a sandbox. Not a virtual machine. Your actual apps and files.

Anthropic built some defenses. Claude asks permission before accessing each app. Some sensitive apps like crypto tools are blocked by default. Users can block specific apps. The system scans for prompt injection attacks.

But the company admits: "these guardrails are part of how Claude is trained and instructed, but they aren't absolute."

The help center warns: don't use computer use for financial accounts, legal documents, or medical information. That's a big limitation.

For enterprise users, there's a bigger gap. Audit logs don't capture Cowork activity. Organizations subject to regulation have no record of what Claude did on their machines. This could be a dealbreaker for compliance industries.

Why now

Anthropic is racing to ship because of OpenClaw. This open-source framework exploded earlier this year. It proved users wanted AI agents that take real action. The community created dozens of "claws" — derivative tools for autonomous control.

Nvidia entered last week with NemoClaw. Now Anthropic is competing. They're betting their advantages — tighter integration, existing subscribers, consumer-friendly interface — can beat free.

What it means

The AI industry has a new defining tension. Ship fast enough to win. But slow enough to stay safe. Anthropic is building in public. Shipping incomplete features. Banking on user tolerance for a 50% success rate today in exchange for better tomorrow.

The moment a failure isn't minor, everything changes.

Google's Efficiency Trick

Google Research just dropped something important. TurboQuant. A set of advanced quantization algorithms that compress AI models massively without losing accuracy.

Key detail. It can quantize the KV cache to just 3 bits. No training. No fine-tuning. That's a huge deal for efficiency.

The results. Up to 8x performance increase on H100 GPUs. Memory footprint reduced by at least 6x. Zero accuracy loss. This will be presented at ICLR 2026.

What it means

AI efficiency just took a big step forward. Models can run faster. Use less memory. Keep their accuracy. This matters for both inference and vector search.

The Warning

MIT Technology Review just released an eBook. Title: "Are we ready to hand AI agents the keys?"

Expert warning: continuing on the current path is "basically playing Russian roulette with humanity."

This eBook explores real autonomy being given to AI agents. And whether humanity is prepared for the consequences.

What it means

While companies race to ship AI agents, experts are warning about the risks. The tension between speed and safety is real. Someone is going to have to draw a line.

ChatGPT Is Now a Shopping App

OpenAI announced richer shopping in ChatGPT. You can now visually browse products. Compare options side-by-side. Get up-to-date information.

The feature is powered by the Agentic Commerce Protocol. Major retailers including Target, Sephora, Nordstrom, Lowe's, Best Buy, and Walmart have integrated.

You can upload an image and find similar items. Ask ChatGPT to refine your search. It all works inside the chat.

What it means

ChatGPT is becoming a super app. Shopping. Coding. Writing. Now browsing. OpenAI is building the everything app. And Sora was in the way.

The bottom line

Three big shifts today. OpenAI is quitting video to become a super app. Anthropic is giving AI real control over your computer. Google is making AI more efficient.

The message is clear. The AI industry is maturing. The fun experiments are ending. The serious applications are beginning.