The AI intern just showed up in your Mac app

Claude Cowork makes agents feel less like demos and more like daily coworkers

midas report banner

Welcome to today’s Midas Report.
Your one stop shop for everything that matters in AI.

Here is what is unfolding today, January 16, 2026….

Our top story comes down to one of my favorites.

This week Anthropic shipped Cowork as a research preview inside the Claude app on macOS, and it is only for the 100 a month plan.

That pricing and gating are the tell…

They are testing with early adopters because this is the kind of feature that either becomes a daily habit or gets ignored forever.

Cowork is built on Claude Code, but it is dressed for people who do not want to touch a terminal. The pitch is simple. Give it a task, let it handle the busywork, and keep you in the loop.

The substance is in what it actually does. It can organize files into folders, convert file types, shrink a too big PDF, turn 20 JPEGs into one PDF, and generate reports from a pile of stuff. It can also use a browser to help manage emails or gather information, but only after it asks permission for each action.

That permissioned design matters because it is where agents stop being a fun demo and start being usable in real workflows.

Cowork runs in a virtual machine sandbox and you explicitly grant access to specific folders. If you do not grant access, it literally cannot see them.

That is the right direction for trust, even if Anthropic still warns you not to expose sensitive data and suggests a dedicated nonsensitive folder due to risks like prompt injection from sketchy web content.

Zoom out and you can see the strategy. Anthropic is not chasing the loudest model benchmark. It is trying to own the work surface where plugins and ecosystem apps will inevitably attach.

Once users get used to an AI that can take actions, the next question becomes which tools it can connect to and which workflows it can automate.

Claude Cowork is Anthropic quietly admitting what most AI tools have been pretending not to be.

Not a replacement for you. A seatmate.

If you run a team, the move is to start testing agent based workflows now, internally, on low risk tasks. Pick repeatable ops, document wrangling, and lightweight analysis, then see what sticks. If you are building on Claude, pay attention to Cowork’s interaction design. The sticky part is not magic. It is the way it asks, waits, and then does the thing.

sponser separator
separator

🧠 The Download

Higgsfield hits $1.3 billion valuation as AI video enters hypergrowth. With 15 million users and a $200 million annual run rate just nine months in, Higgsfield isn’t just a lightning strike… it’s a sign that generative video is becoming infrastructure for digital marketing.

OpenAI shops for US hardware as it moves down the AI stack. OpenAI is quietly laying the groundwork for devices, robotics, and a domestic supply chain, acting less like an API company and more like a platform operator. This signals a land grab for physical interfaces and ecosystem control.

Menlo warns that AI markets are still molten, not mature. Menlo Ventures calls today’s AI categories “proto markets” where business models are in flux and every foundation model drop can break your roadmap.

midas marketpulse separator

TSMC surged to a record high after blowout earnings, with Goldman Sachs hiking its price target to NT$2600. This isn’t just a chip win… As TSMC’s foundry throughput rises, it pulls the entire silicon stack with it, solidifying hardware as AI’s real moneymaker.

Parloa just landed $350M at a $3B valuation, tripling its worth in months as enterprise buyers chase agentic AI that actually replaces workflows. Meanwhile, Skild AI is being valued at $14B for building robotic brains, signaling VC appetite for embodied intelligence.

X’s ban on crypto reward apps wiped 20% off KAITO, showing how tokens tethered to platform distribution can vaporize overnight. Meanwhile, Galaxy raised $75M for a tokenized CLO, aiming to bring structured credit onchain.

A bit of a versus today… I wanted to make a whole bunch of art styles of the same person.

I tried standard ChatGPT against Nano Banana.

Nano Banana did exactly what I wanted… ChatGPT got the layout okay, but actually giving different art styles… it failed.

Nano Banana Pro

ChatGPT

Here is the full prompt used along with a reference image of my face…

“Create a comprehensive art style showcase featuring a male portrait (reference the subject from the provided image) displayed in 45 different art styles arranged in a grid layout. Each style should be clearly labeled and visually distinct. The styles include:
1. Future City Style - Futuristic urban aesthetic with neon lights and digital elements
2. White Box Art Style - Minimalist white background with clean geometric shapes
3. Traditional Chinese Ink Wash Style - Brush strokes with ink wash technique
4. Rococo Style - Ornate, decorative with pastel colors and intricate details
5. Ethnic Style - Cultural patterns and traditional motifs
6. 3D Animation Style - Rendered with depth and dimensional effects
7. Ink Wash Illustration Style - Watercolor-like ink wash technique
8. Handicraft Style - Textured, handmade appearance with visible brush strokes
9. Wild Art Style - Abstract, expressive, energetic brushwork
10. Sci-Fi Black and White Style - Monochrome with futuristic elements
11. Mechanical Style - Industrial, gears, metal textures
12. Naturalism Style - Realistic, detailed, natural colors
13. Rural Idyllic Style - Peaceful countryside aesthetic with soft colors
14. Abstract Composition Style - Non-representational geometric shapes
15. Light Luxury Style - Elegant, sophisticated with gold accents
16. Sports Style - Dynamic, energetic with bold colors
17. Neon Style - Vibrant neon colors and glowing effects
18. Virtual Reality Style - Digital, holographic, futuristic interface elements
19. Red Art Style - Dominant red color scheme with bold strokes
20. Classical Art Style - Renaissance-inspired with rich colors and details
21. Cartoon Style - Simplified, colorful, playful
22. Realism Style - Photorealistic with high detail
23. Watercolor Style - Soft, flowing watercolor technique
24. Black and White Film Style - Vintage film grain, monochrome
25. Cyberpunk Style - Neon-lit, dystopian, high-tech
26. Abstract Art Style - Non-objective, expressive forms
27. Impressionistic Style - Soft, blurred brushstrokes with visible texture
28. Futurism Style - Dynamic, speed lines, modernist aesthetic
29. Oil Painting Style - Rich, textured oil painting technique
30. Modern Minimalist Style - Clean, simple, geometric
31. Romanticism Style - Emotional, dramatic, rich colors
32. Romantic Oil Painting Style - Romantic themes with oil painting technique
33. Surrealism Style - Dreamlike, impossible imagery, unexpected juxtapositions
34. Decorative Art Style - Ornamental, patterned, decorative elements
35. Understated Realistic Style - Subtle, refined realism
36. Ultra-Modern Style - Cutting-edge, avant-garde, innovative
37. Retro Style - Vintage, nostalgic, 1970s-1990s aesthetic
38. Punk Style - Edgy, rebellious, bold colors and graphics
39. Anime Illustration Style - Japanese anime aesthetic with vibrant colors
40. Fantasy Style - Magical, mythical, imaginative elements
41. Minimalism Style - Extreme simplicity, clean lines
42. Baroque Art Style - Dramatic, ornate, rich colors and details
43. Flat Design Style - Modern, clean, two-dimensional
44. Realistic Oil Painting Style - Detailed oil painting with rich textures
45. Light and Shadow Photography Style - Dramatic lighting with strong contrast

Arrange all 45 styles in a clean, organized grid layout with clear labels for each style. Use a professional, gallery-like presentation. The subject should be recognizable across all styles while maintaining the unique characteristics of each art style. Ensure the layout is balanced and visually appealing in a 4:5 aspect ratio.”

Thanks for reading today’s Midas Report! See you tomorrow.

MIDAS AI