Union Alpha AI Guide: Everything We Know

Julian Goldie (0:00)

A brand new AI model showed up online three days ago. It was free, it scored near the top of the charts for coding, and nobody could actually say who made it. It was called Union Alpha. If you never heard of it, that's the point. It appeared with no company, no name attached to it, no lab, no logo, just a name, and the price of zero. Thousands of people started using it that same afternoon. And for a day and a half, the whole AI world was trying to work out who built it. Was it Google? Was it a Chinese lab? Was it a secret test from OpenAI? 33 hours later, the answer came out, and pretty much everyone who was guessing didn't quite get it. So basically, what happened was this was not a company. Union Alpha was never actually one model. And I'm gonna slow down and give you the full story because there's something inside it that changes how you think about picking AI tools for your business, and it costs nothing to understand, but it's super powerful to understand. So here's exactly what happened. On September 16th, 2026, the afternoon, a new listing appeared on a website called OpenRooter. If you haven't used OpenRooter, it's basically like a shop for AI models. You go there, you pick a model, and you plug it into whatever tool you're using. Hundreds of models in one place, one login. Pretty nice.

Now, this new model, Union Alpha, could test. It could read images. It could use tools, which means it could do jobs instead of just chatting. It could hold 262,000 tokens of context, which is like pretty decent at this point. Not quite at frontier level. And the price before was zero for this.

Where the maker's name usually goes, it just sets stealth. Ten minutes later, a coding tool called OpenCode told its users the model was free for the next week. Twenty-seven minutes after that, OpenRooter's CEO posted a chart with Union Alpha sitting as a gold star near the top of a coding benchmark with no number printed next to him. So people piled in and they started digging. And there's normally a way to find out who made a model. It's a bit like handwriting. Every lab's model leaves little fingerprints in terms of how it counts words and how words errors. Now, when an anonymous model called OxAlpha showed up in August, people matched his fingerprint in three days and proved it was ZEIS GLM 5.3 Flash.

Union Alpha's fingerprint didn't match anything, which is pretty interesting. And then on September 17th, OpenRooter posted the name. Union Alpha is a model called Perito made by a company called Unbiased. And here's why no fingerprint matched, because Perito isn't one AI. It's several AI models running at the same time on every single thing you ask it. Their answers get combined into one answer, and that's what comes back to you. The company says Mix includes both big frontier models and open source ones, so you can send one message and then four-ish models quietly work on it, and you get one reply. That's why nobody could find the lab. There wasn't a lab. Now, there were some people who kind of guessed what was going on here in the background, but it's pretty interesting. So we're going to run through how it's performing, what you can use it for, and how you can get the most out of Union Alpha. Now, something to note here.

Unbiased actually made a point of saying Peruto is not a router. A router is the thing that reads your question, guesses how hard it is, and sends it to one model. Cheap question goes to a cheap model, hard question goes to an expensive one, sounds sensible. Unbiased says if a reliable way to do that existed, they'd use it.

That is something that JEV. AI I've seen is possible. But with this system, basically when you talk to an AI agent, the beginning of the conversation gets sent again every single time, right? Your instructions, your rules, your background files, all over and over and every step. Caching means the provider remembers that chunk, so you get a discount on it. If you look at the prices they've actually published, $2.50 per million tokens in, per million words in, roughly, cached input is 25 cents, and 10 for the price. Now, here's the problem with routers. If the system swaps models halfway through, that memory discount is actually gone, so you go back to the full price and everything you already sent. Now, I run an agent setup where Claude Code, Hermes, and OpenClaw all share one memory, right? Long jobs, same instructions, loaded every stop.