Ultramarine AI Is INSANE! 60X Cheaper Than Claude?!

Julian Goldie (0:00)

Ultramarine AI is about 60 times cheaper than Claude for office work, and on one big test, it scored higher. So let me run you through the numbers and how this works. A company called Paper Instruments gave this new model a big stack of word tasks, client reports, spreadsheets, slide decks, edits to real-world files. And Ultramarine OneFlash did each task for 13 cents. Claude FableFive did the same task for $8.19 each. Ultramarine also scored 59.3. FableFive scored 57.8. And the team that trained it was just five people. Their CEO, Dhanash Kasi, says the whole training actually cost $300,000 all in. For a model in the same range as Claude and GPT, that's a tiny team and a tiny budget. It's also open source, so you can use it through their API right now. The model files themselves, the weights, the thing you download to run it on your own machine are coming in about two weeks. So why should you care if you never write a line of code? Because this model wasn't built to code. It was built to do the kind of work most of us do all week. Reports, contracts, forecasts, memos for the boss, and the way they trained it is the real lesson here. I'm calling it the judgment machine. By the end of this, you know how it works, why it made a cheap model this good, and how the same idea makes any AI agent you use better. That could be Claude, could be Hermes, could be OpenClaw, could be any of them. Plus, I'll show you where Ultramarine is weak, because it is weak in one area. You need to know it before you switch anything over. And let's start with one line from the paper. The team actually wrote that most professional work is not verifiable. Put simply, mostly work has no answer key. So put simply, your work or most of it has no answer key. And if you hold on to that, everything in this model basically comes back to it. Now, for the last couple of years, AI has gotten really good at coding. Why coding? Because code has an answer key. You run the code, it works or it breaks. A computer can grade it in one second, and labs train AI with a method called reinforcement learning. It can explain words, it's practice, plus a score. The AI tries a task, it gets a score, it does more of what's scored well. Not only works if you can score the task, and code is very easy to score. The paper says coding has held the biggest share of this practice time across the industry. Now, think about a monthly SEO report for a client. Is it good? There's actually no pass or fail. It depends on judgment. Did it say which page is moved up? Did it explain why? Did it say what happens next month? Did it hide the bad news at the bottom? Two smart people could write very different reports. Both could be great. That's what non-verifiable means. Worker or computer can't just check the answer. And that's most work, right? Sales school notes, proposals, a picture of a new client, a plan for the next quarter. So AI gets lots of practice on the work of an answer key and way less practice on the work you actually do all day. The Ultramarine actually goes straight at that gap. So what did they do that's different? Well, three things. And the third one is the one that I would copy first. The first is how they grade. So when there's no answer key, you need a rubric. A rubric is a checklist that says what good looks like. Each line is worth points. Easy, right? Except most rubrics are bad, and the paper explains why in a way I haven't seen. Anyone else explain this clearly. So every line in the rubric joins and draws a line between answers, passes and answers that fail. If you draw it too tight, you fail good answers. There example, a line that says the answer must use the exact phrase, sole forecast basis. Two perfect answers said the same thing in different words, both failed. Now, if you draw it too loose, you actually pass junk. A line like provides a useful update for the board that passed answers that only said half what actually matters. The fix is a line that tests one clear idea and names the facts needed to check it. Picture, for example, like grading a book report. You could put uses the word because in sentence two is too tight, writes a good report is too loose, give one reason the dog ran away using a fad from the story. It's just about right. So the team actually said reliable grading comes down to how each line is drawn, more than picking a smarter AI to do the grading. And then each line gets a yes or no. The lines get weights based on how much they matter to a real professional. If you add it up, you get a score for work that has no answer key. The second thing they actually did differently is where the practice happens. They built fake offices, 88 kinds across four areas, finance, law, consulting and operations, split into 19 areas. Each fake office has files, spreadsheets, contracts, emails, filings and the live system that AI can actually change like a sales CRM or a company's books. Then they mix things up. Who's the client? Who's the format? Which two documents disagree with each other? What stage is the deal at?