A mysterious new model just took over the internet... Transcript, AI Summary & Key Points

00:00 A few weeks ago, while I was busy putting the final touches on my Dolly Parton is too fast for the Grim Reaper tattoo, someone anonymously dropped a new frontier grade model online. They called it Ox Alpha. And in its first 6 days, it served 42 trillion tokens, which easily made it the most popular model on Open Router, doubling the next closest, which was DeepS.

00:17 It comes with a million token context window, accepts video input, writes code like it has trouble making eye contact, and until just a few days ago, it was completely free. as some assumed it was Jepu, others Xiai, and at least one guy was convinced that Google actually shipped something useful. In today's video, we'll look at how this anonymous model won over the internet, who's behind it, and see if it can finally help me realize my horse SAS dream.

00:40 It is September 1st, 2026, and you're watching the Code Report. It hasn't even been 2 weeks since we talked about how Silicon Valley is terrified of what's happening in China after the release of Deep Seek Harness. And unfortunately, Daario and Sam's terrible, horrible, no good, very bad month may have gotten even worse because this new anonymous model had made in China written all over it.

00:59 And as we know, things that escape out of Chinese labs tend to have a big impact on the rest of the world. The first clue was that Ox Alpha was the fifth anonymous animal model to show up on Open Router recently, and the four before it all turned out to be from China. And even more convincing, some devs noticed that its stack traces and error codes matched Zoo's API, and a tokenizer fingerprint test came back matching their GLM series, but nobody waited for a confirmation to start using it.

01:26 Yes, the fine print did say that all your prompts would be retained by a mysterious, probably foreign entity. But nobody actually cares about privacy when the magic box is free and works this well. So devs happily pasted their company's proprietary code into it, routed claude code through it to dodge their rate limits, and at its peak, OX Alpha accounted for nearly a third of Open Router's entire weekly traffic.

01:45 The hype machine got so out of control that a misleading screenshot showing an 80% score on the deepsw S.WE benchmark went viral, even though the real number turned out to be closer to 58, and someone even speculated it was so good it might be a stolen Claude checkpoint. Sure enough, on August 26th, Jepu came clean and announced that Ox Alpha was GLM 5.3 Flash, a natively multimodal mixture of experts model with 320 billion total parameters and its MIT licensed weights hit Hugging Face the same day.

02:15 But the detail that really killed the vibe in Silicon Valley was that the entire Stealth run was served on just 100,000 Chinese-made chips, which I'm told really hurt Jensen's feelings. Of course, the reveal also meant the free samples were over and the anonymous endpoint was replaced with an actual pricing page, but it's still stupid cheap at just 15 cents per million input tokens and 50 cents per million out.

02:37 And with a 50% discount through September 9th, it's currently up to 40 times cheaper than Claude. And as we know, normal launches like this come with a bunch of Trust Me Bro benchmarks that usually fall apart the second anyone touches them. But their benchmarks hold up pretty well under outside scrutiny. The catch is that it's slow, loves to yap, and occasionally gets stuck in doom loops.

02:56 But in my opinion, there's only one way to test if it's actually any good, and that's by using it on a dear project of mine. After failing to raise funding for Horse Tinder, I've pivoted to its once beloved cousin, Horse, a 2016 Angular.JS app whose last commit was 10 years ago. So, I asked it to modernize the legacy stack without losing the humor and vibes I had so carefully crafted.

03:17 It correctly detected the outdated technologies and then proceeded to narrate its plan of attack in every little detail for at least 20 minutes. It then started building a design system in raw dog CSS. It went with a beautiful purple slot background and rainbow gradient buttons showing true inspiration, sorry, distillation from other popular LLM's visual tastes.

03:38 Because of its vision capability, it was able to detect a CSS overflow bug on mobile screens, properly diagnose it, and fix it with the cursed min width equals zero trick that always seemed to work. It added some nice enter animations for video cards, which was a really nice touch. All the bits of humor are still there, and it even added some new ones like these toast notifications.

03:59 The code-wise, it replaced the legacy stack with no stack. It wrote the app in HTML, CSS, and vanilla JavaScript and wrote code that looks like a legacy codebase of its own. But I really wanted to test its vision skills. So I fed it a random horse video and asked it to add it to the app and create a title, description, and comments that are relevant to what's in the video.

04:16 It decided to use FFmpeg to break down the video into a series of frames, analyze those frames, and it did a really good job. It got the full context of what happens in the video, nailed the description, and even align the humor with the rest of the platform. And another tool worth checking out is Exa, the sponsor of today's video. It's a search engine built for AI agents that gives you a lot more than just web results.

04:38 It includes 350 million research papers, financial data, and info on over 1 billion people and 70 million companies with new sources added every week. I hooked it up to Polyhip to let my users gamble on the stock market in even more degenerate ways, like betting on which tech CEOs will sell company stock right before their next earnings call. And EXA was able to pull the recent SEC filings and convert it into structured JSON in one shot.

05:02 It also cites all of its sources, so users can't accuse me of cheating anymore. You can plug directly into Claude, Chat GPT, or your other favorite agent using their native plugins. Or try Exa Agent to get answers to more complex queries that need to pull from multiple data sets simultaneously, which is a lot more powerful than what LLM can do on their own. Try Exa today for free and get $20 in credits at the link below. This has been the Code Report. Thanks for watching, and I will see you in the next one.