A few days ago, a new AI model quietly appeared on OpenRouter.
Its name is Ox Alpha.
Nobody officially knows who made it.
It is listed as stealth/ox-alpha, described as a reasoning model built for coding and sustained agentic work, and it comes with some pretty wild specs, including a 1,048,576-token context window and up to 131,072 output tokens. The developer is simply listed as an anonymous third-party provider.
Rather than throw benchmarks at it, I did what I usually do when I want to get a feel for a coding model.
I told it to make me something.
So I Had Ox Alpha Build a Game
I gave Ox Alpha a game-building task and ended up with Pizza Rush, a little pizza-shop game where customers place orders and you have to build, bake, and serve their pizzas before they lose patience.
You can actually play it here:
The game has more going on than I expected from a quick model test.
Customers request pizzas. You start with dough, add ingredients by dragging or tapping them, make sure you remember the sauce and cheese, bake the pizza, and try to pull it from the INFERNO 9000 oven while the oven meter is in the green zone.
Meanwhile, customer patience is ticking down.
Three customers walk out and your pizza empire falls apart.
It also tracks days survived, pizzas served, money earned, and perfect pizzas.
What interested me wasn't that the model could generate HTML, CSS, and JavaScript for a game. Plenty of coding models can do that now.
It was that it assembled several systems into something that felt like an actual small game instead of a decorated page pretending to be one.
There was state.
There were orders.
There was timing.
There were failure conditions.
There was scoring and progression.
There were multiple mechanics interacting with one another.
It was a surprisingly solid first impression.
And then I started wondering what the hell I was actually using.
Nobody Knows Who Made Ox Alpha
Ox Alpha was released as a stealth model, and the provider has deliberately chosen to remain anonymous during the preview.
That immediately turned its release into an internet detective story.
Business Insider reported that developers were already impressed by its coding and long-running agent capabilities, while speculation around its creator bounced between several companies, including Chinese AI lab Z.ai and Microsoft's MAI team.
TechCrunch reported the same uncertainty, with theories changing almost as quickly as people could test the thing.
But the most interesting part of the story isn't simply that nobody knows who made it.
It's how people are trying to figure it out anyway.
You Can Hide a Model's Name. Hiding Its Fingerprints Is Harder.
Researchers started treating Ox Alpha like a black box.
Instead of asking:
"Who made you?"
they started examining characteristics the model cannot easily disguise.
Its tokenizer.
Its reasoning controls.
Its API parameters.
Its error messages.
Its output behavior.
Its tool-calling behavior.
Basically, they started fingerprinting the model.
One public investigation compared Ox Alpha against Z.ai's GLM-5.3 and Xiaomi's MiMo-V2.5 using multiple kinds of black-box testing.
The strongest result came from tokenization.
Across 24 baseline-adjusted tokenizer probes, Ox Alpha matched GLM-5.3 on all 24.
Even stranger, Ox Alpha consistently reported exactly 75 additional prompt tokens on every test, which is consistent with the two systems sharing a tokenizer while Ox Alpha adds a fixed hidden system prompt or template overhead.
MiMo only matched five of the 24 probes.
That doesn't prove Ox Alpha is GLM-5.3.
A compatible tokenizer could also appear in a related model, fine-tune, unreleased checkpoint, distillation, adapter, or intentionally compatible serving system.
But it is considerably stronger evidence than simply saying:
"It feels like GLM."
Then the Other Clues Started Matching Too
The tokenizer isn't the only thing pointing in that direction.
Researchers also found that Ox Alpha and GLM-5.3 expose very similar mandatory reasoning behavior and parameter capabilities.
Attempts to disable reasoning produced matching or highly similar provider behavior.
Deterministic outputs also showed unusually close formatting and response patterns.
For now, what model this actually is still a mystery.
Just one with fingerprints all over the crime scene.
Model Fingerprinting Might Become a Whole Thing
This is the part I find most fascinating.
We usually think of a model's identity as the name attached to the API endpoint.
GPT-whatever.
Claude-whatever.
Gemini-whatever.
GLM-whatever.
But a model also has observable characteristics.
A tokenizer can be a fingerprint.
Error formatting can be a fingerprint.
Parameter validation can be a fingerprint.
Reasoning controls can be a fingerprint.
Tool-call formatting can be a fingerprint.
Unicode handling can be a fingerprint.
Even the number of tokens produced by carefully chosen strings can reveal information about the model family underneath.
That means an anonymous provider can hide the label without necessarily hiding the machinery.
With enough probes, you can start narrowing down what you're talking to without ever seeing the weights, training pipeline, or internal architecture.
It reminds me a little of browser fingerprinting.
Your browser doesn't have to explicitly tell a website exactly what it is if enough tiny behavioral characteristics collectively identify it anyway.
Apparently AI models can have the same problem.
There Is Another Interesting Experiment Happening Here Too
The anonymity itself may actually be useful.
Think about how much baggage normally comes with a model launch.
If OpenAI releases something, everyone compares it to Claude.
If Anthropic releases something, everyone compares it to GPT.
If Google releases something, half the conversation becomes Gemini versus everyone else before people have even finished testing it.
Brand names affect expectations.
Benchmarks affect expectations.
Pricing affects expectations.
Previous experiences with a company's models affect expectations.
Now imagine releasing:
model-a
model-b
model-c
No company names.
No leaderboard announcement.
No giant launch presentation telling people how incredible they are supposed to think it is.
Just put the model in developers' hands and see what they actually do with it.
Do they keep using it?
Do agents successfully complete tasks?
Where does it break?
What kinds of projects do people trust it with?
What happens when developers aren't evaluating the company logo alongside the output?
Ox Alpha is currently giving its creator a pretty fascinating real-world experiment, whether that was the intention or not.
My Own Test Doesn't Tell Me What Ox Alpha Is
Pizza Rush obviously isn't a scientific benchmark.
I'm not going to claim that one little pizza game proves Ox Alpha beats GPT, Claude, Gemini, GLM, or anything else.
It doesn't.
What it gave me was something different.
A reason to keep testing it.
My first experience with Ox Alpha was a model taking a fairly open-ended coding task and turning it into a coherent, playable little game with multiple interacting systems.
That made me curious.
The mystery surrounding it made me even more curious.
And then discovering that people were identifying its possible origins through tokenizer differences, API quirks, error messages, and black-box testing sent me straight down the rabbit hole.
At this point, I'm almost less interested in whether Ox Alpha turns out to be the best coding model.
I want to know whether the internet managed to identify it correctly before its creator ever said a word.
Because if the GLM theory eventually gets confirmed, there is a pretty interesting lesson buried in this whole experiment:
You can make an AI model anonymous.
Making it unidentifiable might be a lot harder.
Top comments (1)
i try the game pretty neat ^_-