03-06-Daily AI News Daily

Today’s Summary

GPT-5.4 lands desktop control with 75% success rate beating humans, million-token context API pricing is actually reasonable.
Raycast generates Mac apps with natural language, key Alibaba Tongyi figure departs, CEO steps in personally.
Desktop Agent era kicks off, developers should try each one today.

⚡ Quick Navigation

💡 Tip: Want to experience the latest AI models mentioned here (Claude 4.5, GPT, Gemini 3 Pro) right now? No account? Grab one at Aivora — one minute setup, hassle-free support.

Today’s AI News

👀 One-Liner

OpenAI dropped GPT-5.4 last night—it can control your desktop directly, with success rates higher than humans.

🔑 3 Keywords

#GPT5.4Explosion #AIDesktopControl #AlibabaLeadershipShakeup

🔥 Top 10 Headlines

1. OpenAI Releases GPT-5.4: First General-Purpose Model That Can “Operate Your Computer Hands-On”

Picture this: you tell AI “organize this Excel sheet into a PowerPoint and email it out,” and it actually opens Excel, drags things around, switches to PowerPoint, formats it, opens your email—all you do is watch. That’s GPT-5.4. In desktop control tests, it hit 75% success rate, straight-up beating humans at 72.4%. Previous GPT-5.2 was only at 47.3%. Investment banking modeling jumped from 68.4% to 87.3%, and coding ability absorbed the entire Codex specialized model. The real win is “tool search”—token consumption dropped 47%, so developers’ wallets can finally breathe. Plus users get access today. The wait-and-see crowd wins again.

Image

2. GPT-5.4 Opens Million-Token Context, Double Pricing Only for Overage

The old pain point with long documents was always context window limits—can’t fit it all, so you’d slice, splice, and shuffle back and forth. Now GPT-5.4’s API and Codex support 1 million tokens natively. Here’s the kicker on pricing: up to 272K tokens charge normally, only the overage gets double-charged. Compared to Claude’s million-token pricing, this is genuinely better value. Code ability is now baked into the main model—no need to call Codex separately anymore, one model does it all. Developers are thrilled, wallets are smiling.

Image

3. Are GPT-4.5 and o1 Pro the Real Peak? Community Debates Whether 5.2 Is “Budget Tier”

Just as everyone’s celebrating GPT-5.4, the community’s stirring up a nostalgia debate. Some users pulled test data showing that outside of coding Agent scenarios, GPT-4.5 and o1 pro still reign supreme in overall quality. Others straight-up said “GPT o3 beats 5.2 thinking, 5.2 is a scam.” Harsh words, but they reflect reality—version numbers go up, actual experience doesn’t always follow. Now that 5.4 dropped, is it a true all-around leap, or just another “strong here, weak there” tradeoff? Worth testing yourself.

4. Alibaba Confirms Tongyi Qianwen Lead Lin Junyang Departs, Who’s Really the Soul of Qwen?

This news is peak drama. Alibaba officially confirmed that Lin Junyang, core figure at Tongyi Qianwen, has left. CEO Wu Yongming personally heads up a new foundation model support group. But here’s the juicy PR move—Alibaba’s comms team quickly reframed it: “Lin wasn’t a core person, just active on overseas social media,” claiming “Qwen’s soul is Alibaba Cloud’s CTO.” In the internal memo, everyone else gets called by nicknames, but Lin gets his full name. Netizens: “They use nicknames before the breakup, full names after.” Big tech drama with major implications—Alibaba’s open-source LLM strategy is worth watching closely.

Image

5. Raycast Launches Glaze: Generate Mac Apps with Natural Language

Building a simple Mac tool used to mean learning Swift, wrestling with Xcode, dealing with code signing certificates—just the setup would scare off half the people. Now Raycast dropped Glaze: describe what you want in plain English like “I want a countdown timer” or “make me a Markdown editor,” and it generates a native Mac app—you can even publish to the App Store. The barrier to custom software just hit the floor. This isn’t demo-level toy stuff, it’s a real product. For creative people who can’t code Swift, this door just opened wide.

6. Claude Code Remote Control Magic: Interact via Feishu, Telegram, Discord Anytime

In a meeting across town and suddenly need Claude Code to fix some code—used to mean waiting to get back to your desk. Now this open-source Skill plugs Claude Code into Feishu, Telegram, Discord so you can remote-command from your phone, approve tool calls, watch output in real-time. Setup is an interactive wizard that walks you through “click here, fill that”—even non-techies can handle it. Keys stored with chmod 600, logs auto-sanitized, security details are solid. One-line install: npx skills add op7418/Claude-to-IM-skill. Worth trying.

Image

7. TuriX-CUA Open Source: Let AI Control Mac and Windows Desktop Like Humans

Right on cue with GPT-5.4’s Computer Use, the open-source community dropped their own answer the same day. TuriX-CUA is a desktop control Agent framework that lets AI see the screen, click the mouse, type—book flights, search YouTube and like videos, grab files from Discord, make charts, reply to your boss, all doable. Architecture splits AI into “brain, executor, planner, memory manager”—each role can swap different models, super flexible. Already has Skills to plug into Claude Code. Doesn’t need target software to provide APIs, if you can click it, it can click it.

8. Codex Desktop Finally Lands on Windows

Windows users waited long enough. OpenAI’s Codex desktop app was Mac-only forever, today it finally hit Windows. Paired with GPT-5.4’s launch, Windows developers can now enjoy million-token context, native code ability, and Computer Use features right on desktop. Full feature rollout details are still coming, but just “no more jealous of Mac users” is enough to get the community hyped. Download it now and see how much your dev workflow speeds up.

9. Apple Releases M5 Chip Series: AI Performance Quadruples, MacBook Pro Battery Lasts 24 Hours

That charger you bring on business trips might actually be useless now. Apple’s new M5 Max chip quadruples performance on on-device AI tasks, new MacBook Pro runs 24 hours straight on a full charge—full day of meetings, coding, running local models, no outlet hunting. Studio Display XDR upgraded too, 5K resolution + 120Hz refresh. Apple being Apple, the budget line is still “heartbreaking” (iPhone 17e still 60Hz notch), but the high end really delivered this time. If you run local AI models, M5 Max is worth your attention.

10. Google NotebookLM Launches “Cinema-Grade Video Overview”: Study Notes Become Movies

What’s the ultimate form of organized study notes? Mind maps? PowerPoint? Google says nope, we’re making movies. NotebookLM’s new video overview feature auto-generates cinema-style explainer videos from your uploaded materials—narrative structure, visual style, pacing all handled. Multiple AI models working together behind the scenes, script to visuals fully automated. Currently only for Google AI Ultra paid users, English only. High barrier, but the direction is spot-on: future learning might actually be “watching movies.”

📌 Worth Watching

😄 AI Fun

Manus Says It’s One Year Old, But… It’s Only Been a Few Months? 😂

Manus posted a celebration tweet today saying “🎂Manus turns one today,” but netizens instantly called it out: you literally just came out this year? Turns out AI Agents don’t just help you work, they’ve learned to lie about their age too. Even Baozong couldn’t resist commenting: “Feels a bit hallucinated.” AI hallucination problem, this time turned on itself.

Image

🔮 AI Trend Predictions

GPT-5.4’s Computer Use Ability Ignites Desktop Agent Ecosystem

  • Prediction Timeline: April-May 2026
  • Confidence: 80%
  • Reasoning: Today’s news of GPT-5.4 native desktop control + TuriX-CUA open source launching same day shows desktop control Agent infrastructure maturing fast. Expect wave of vertical applications built on Computer Use within two months.

Natural Language App Generation Becomes New Track

  • Prediction Timeline: Q2 2026
  • Confidence: 70%
  • Reasoning: Today’s Raycast Glaze launch generating Mac apps via conversation, plus prior similar products, shows “talk to build apps” shifting from concept to product.

Alibaba Qwen Team Releases Major Update Soon to Stabilize Morale

  • Prediction Timeline: March-April 2026
  • Confidence: 65%
  • Reasoning: Today’s Alibaba confirms Lin Junyang departure , CEO personally leading new group. Big tech typically accelerates releases after core team changes to answer market concerns.

OpenAI Rolls Out More Agent Tool Integrations Before GPT-5.2 Retirement

  • Prediction Timeline: April-June 2026
  • Confidence: 75%
  • Reasoning: Today’s GPT-5.4 revealed Tool Search feature and Codex Windows version , plus 5.2 retiring June 5, OpenAI will densely ship Agent ecosystem tools during transition.

❓ Related Questions

How to Experience GPT-5.4?

GPT-5.4 currently requires ChatGPT Plus, Team, or Pro subscription to use, API access also needs paid accounts. Domestic users may face payment difficulties or account registration restrictions.

Solution: Visit Aivora to get ready-made accounts, instant delivery, worry-free support.

Last updated on
03-07-Daily03-05-Daily