DeepSeek just undercut everyone again 🐋

Good morning. DeepSeek dropped a new model that's so cheap to run it makes everything else look overpriced, which is exactly how it shook up the whole industry last year. NVIDIA built a tool that scans a video and tells you if it's AI in the blink of an eye, aimed at newsrooms drowning in fakes. And there's a storytelling trend all over Instagram and TikTok right now that uses AI photos to show how far your life drifted from the plan. What tool are you running your boring back-end tasks through these days? Hit reply and tell me. We cover all three below.

DeepSeek put out a new model called V4-Flash on July 31, and an independent research firm just crowned it the cheapest well-known AI model in the world to actually run. Not a little cheaper. It came in at about 3 cents to complete their full test, versus around 3 dollars for Claude Fable 5 doing the same work. That's the gap that made DeepSeek a household name last year, when its first model tanked tech stocks and made everyone ask why US AI was costing so much.

The number that matters isn't the sticker price, it's the cost to finish a real job. A model can look cheap per word and still run up a fat bill if it rambles for ten paragraphs to answer a simple question. Not the sharpest model out there, but plenty capable for most everyday work, at a price that rounds to nothing.

This is the thing to watch even if you never touch DeepSeek yourself. When the floor drops this low, everybody else has to follow, and the price of the AI baked into the tools you already use keeps sliding down with it. If you run any kind of automation, a model this cheap means you can point it at high-volume grunt work, captioning, sorting, rewriting, tagging, generating a hundred title options, without watching a meter tick. The catch is it's Chinese and open, so for client work or anything sensitive, know where your data goes before you wire it in.

NVIDIA showed off a new tool called Synthetic Video Detector that looks at a clip and scores how likely it is to be AI-generated, and it does it fast, around 22 milliseconds for a full HD video. It's built for newsrooms and broadcasters, the people who need to know if that "leaked" clip is real before they put it on air. It goes frame by frame, hunts for the fingerprints AI leaves behind, and hands back a probability instead of a yes or no.

The honest part is where it struggles, and it's the part that matters most for people like you. On a clean, uncompressed file it's right about 92% of the time. But every platform, TikTok, Instagram, YouTube, crushes your video with compression the second you upload, and that compression scrubs away the exact clues the detector leans on. At heavy compression it drops to around 82%, meaning it's wrong on roughly one clip in five, right where the sketchy footage actually lives. The detector is weakest in the same spot the fakes are strongest.

For creators this cuts two ways. It's a preview of a world where platforms and viewers run everything through a checker, so anything you make that even smells synthetic could get auto-flagged, fairly or not. But it's also proof that catching AI is genuinely hard and getting harder, which is why "it looked real" stopped being a defense. The takeaway from the last few weeks holds: don't build your reputation on passing fakes off as real, because the detection net is tightening, and being the creator who got caught is a worse look than just labeling your work.

There's a format eating Instagram and TikTok right now that goes by a few names, Plot Twist, Expectation vs Reality, Plan A vs Plan B. The setup is always the same. You show the life you carefully planned, the city, the job, the person, the timeline, and then hit the "plot twist" turn where you reveal how completely different it actually turned out. It works because it's the most relatable thing in the world. Nobody's life went according to the plan.

The AI part is what makes it clickable. Instead of digging through old camera rolls for photos that match the story, people are generating the images: the wedding that didn't happen, the city they almost moved to, the version of themselves that took the other job. Some run their real photos through AI to restyle them into a clean matching set, some generate the whole thing from scratch. It turns a story you could only tell with a wall of mismatched screenshots into a polished little film.