Union Alpha: I Tested the Free Stealth AI Model
Union Alpha: I Tested the Free Stealth AI Model
Union Alpha is a free stealth AI model available right now through OpenRouter, and it's posting benchmarks that go toe-to-toe with the best models from OpenAI and Anthropic. Nobody knows which lab built it. It has a 262,000-token context window, costs nothing to use today, and in my testing it held its own against Opus 5, Fable, and Astra across front-end design, game building, motion graphics, and long-context recall. Here's exactly what I found when I put it through the wringer.
What Is Union Alpha?
Union Alpha is an unreleased "stealth" model being served anonymously through providers like OpenRouter. A stealth model is one an AI lab pushes out under a codename before an official launch — usually to gather real-world usage data and benchmark feedback without attaching its brand. So we genuinely have no idea who's behind this one yet.
Here's what we do know:
- Context window: 262,000 tokens
- Price: completely free right now
- Access: through OpenRouter, paired with a harness like OpenCode
That 262K context window is the headline spec. It puts Union Alpha in the same tier as the big frontier models for how much text it can hold in memory at once, which matters a lot for real coding work where you're feeding it entire files and long histories.
How Good Are Union Alpha's Benchmarks?
The early benchmark numbers are strong — with the usual caveat that they're unofficial and worth a healthy dose of skepticism.
The DeepSWE score, posted by the CEO of OpenRouter, shows Union Alpha performing at the same level as Astra, Fable, and Opus 5 — but at a fraction of the cost. That's the wild part. If the pricing ends up matching what that chart implies, you're getting frontier-level output for pocket change.
There were more stats too: Union Alpha reportedly outperforms Opus 5 on Terminal Bench 2.1, along with strong showings on SWE-bench Verified.
Now, a real caveat here. Benchmarks are getting harder to trust. When you see something like Gemini 3.8 Flash scoring absurdly high on the same charts, you have to ask whether any single benchmark is the be-all and end-all. The answer is no — no benchmark is. But if these numbers hold up even directionally, that's a great sign for a free model from an unknown lab.
The bottom line: treat the benchmarks as a reason to test the model yourself, not as gospel. That's exactly why I ran my own head-to-head tests.
How Do You Get Access to Union Alpha?
Getting Union Alpha running takes two pieces: a provider and a harness. Here's the setup I used.
- Create an OpenRouter account. Go to openrouter.ai and sign up. It's free, and using Union Alpha through it currently costs nothing.
- Connect a harness. You need something to actually code with. I used OpenCode, which connects to OpenRouter in just a few clicks.
- Select the model. Once OpenRouter is wired into OpenCode, open the model picker and choose Union Alpha.
That's it. No API tier gymnastics, no waitlist. One thing to note: Union Alpha has no effort settings. Models like Opus 5 let you dial reasoning effort up or down — Union Alpha just runs at one setting, so every comparison below is Union Alpha's default against a competitor's high-effort mode.
How Does Union Alpha Handle Front-End Design?
I ran two front-end tests — one bare-bones prompt and one with more direction — to see how the model behaves with and without guardrails.
Test 1 — the minimal prompt. I asked it to "create me a landing page for a fake boutique hotel in London" and explicitly told it not to use any skills. In 3 minutes and 37 seconds, it produced a clean hero section with a rolling banner. What I liked most: it didn't scream AI. No default rounded-corner cards, none of the standardized layouts you see everywhere. It went a little off the beaten path, and the color scheme was solid. Simple? Sure — but that's a function of a simple prompt, and it knocked it out fast.
I gave Opus 5 the exact same prompt on high effort. Opus edged it out. Opus added a smart "check availability" button up top that jumped you to the booking section, went deeper on the background with what looked like SVGs, and threw in a bit of scroll motion. I gave the slight edge to Opus — but here's the key point: losing to Opus doesn't make Union Alpha "worse" when you factor in cost. Until we know the real price, a near-Opus output from a free model is a win.
Test 2 — the detailed prompt. This time I asked for "a polished, responsive homepage for Dune House, a fictional boutique desert hotel in Joshua Tree, California," with some guardrails but still room for the model's own judgment on layout. The jump was huge. The hero image alone elevated the whole page, and the result was far more polished than the first test.
I ran the same prompt through Astra. Both went for a full-bleed hero image, both used images well as you scrolled. Astra had a cleaner availability section, so I'd give it a slight edge — but the gap was tiny. We're talking a couple of tweaks away from being the same page. For a free stealth model to land that close to Astra on a real design brief is a big thumbs up.
Can Union Alpha Build a 3D Browser Game?
This test got messy, and it's worth being honest about that.
I first tried the exact Fortnite-clone prompt (with reference images) I'd used in a previous Fable-versus-Astra video. Union Alpha errored out every single time. I can't say for sure whether that was an OpenRouter routing issue or a genuine capability limit — it's possible I simply gave it too much.
So I pivoted to a 3D arcade racing game for the browser on a stylized coastal highway, with guardrails on what to build. It still hit some errors mid-build and I had to restart the session — again, unclear whether that's the model or the routing.
The final result, built with Three.js, actually held up. The controls worked, it ran smoothly, there was real physics when you hit things, and the graphics weren't bad — this was a single circular track, nothing complicated, but it played nicely. I ran the same prompt through Astra and got a strikingly similar result: same tier, smooth handling, working physics, comparable graphics. When Union Alpha matches a top model on a task like this, I count it as a win — even with the rough edges getting there.
How Does Union Alpha Do With Motion Graphics?
For the motion test, I told it to use the Higgsfield motion skill to create a 2D motion explainer about how a message travels across the internet — packets, undersea cables, thousands of miles in half a second.
The output was genuinely solid, and so was Astra's version of the same brief. The important context: this test rides on the back of that motion skill, which is really just calling an MCP server to generate the clips using Seedance under the hood. So it's less a test of raw model intelligence and more a test of whether the model can pick up a skill and use it correctly. The takeaway: throw the right skills at Union Alpha and it operates just like a frontier model.
How Well Does Union Alpha Handle Long Context?
This is where Union Alpha really shined. I ran the 8-needle test — a stress test for long-context recall.
Here's how it works. You flood the model's context window with text, and buried inside are eight similar-but-distinct passages — imagine eight slightly different poems all titled "a poem about flamingos." At the end you ask the model to return exactly the second one. To succeed, the model has to identify what counts as a flamingo poem, separate all eight, and pull back the precise one you asked for.
The results:
- 128K version: 100% — a perfect score.
- Modified 200K version: 99.6% — off by just five apostrophes.
Filling nearly the entire context window and still getting near-perfect recall is exactly what you want from a model with a 262K window. A big context window is useless if the model loses track of what's inside it — Union Alpha doesn't.
Is Union Alpha Worth Using?
Across every test — front-end design, game building, motion graphics, and long-context recall — Union Alpha performed at roughly the same level as Opus 5, Fable, and Astra. It lost a few by a hair, matched the rest, and aced the long-context test.
The whole story hinges on one unknown: price. If Union Alpha launches at anything close to the cost implied by that DeepSWE chart, it's a model most of us should seriously consider adding to our stack. Frontier-level output at a fraction of the price is the kind of thing that reshapes how you pick a model for a given job. For now, it's free — so there's zero reason not to test it yourself while you can.
Keep an eye on this one. When the lab behind it finally reveals itself and drops official pricing, it could shake up the field.
Frequently Asked Questions
What is a stealth AI model?
A stealth model is an unreleased AI model that a lab serves anonymously under a codename, usually through a provider like OpenRouter. Labs do this to collect real-world usage and benchmark data before an official, branded launch. Union Alpha is a current example — nobody publicly knows which lab built it yet.
Is Union Alpha free to use?
Yes. As of now, Union Alpha is completely free to use through OpenRouter. You just need a free OpenRouter account and a harness like OpenCode to run it. That could change once the model officially launches and pricing is announced.
What is Union Alpha's context window?
Union Alpha has a 262,000-token context window. In my long-context testing it scored 100% on a 128K 8-needle test and 99.6% on a modified 200K version, so it holds up well even when the window is nearly full.
How does Union Alpha compare to Opus 5?
In my testing, Opus 5 held a slight edge on some front-end design tasks, but Union Alpha matched or came very close on most others and reportedly outperforms Opus 5 on Terminal Bench 2.1. Given that Union Alpha is free and Opus 5 is not, the value proposition tilts hard toward Union Alpha if the eventual pricing is competitive.
How do I access Union Alpha?
Create a free account at openrouter.ai, connect a coding harness like OpenCode to OpenRouter, then select Union Alpha in the model picker. There's no waitlist or API tier requirement, and it's currently free to run.
If you want to go deeper into testing and building with frontier AI models, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.
