Ox Alpha: The Free Stealth AI Model Nobody Will

There's a model on OpenRouter right now that nobody will claim. It's called Ox Alpha, it's completely free, it has a 1 million token context window, it handles text, images, and video—and if you ask who built it, you get silence.

That's either a clever marketing play or something more interesting. Possibly both.

The launch announcement had the energy of a deliberate event. OpenCode tweeted it first: stealth model, free for a week, 1 million context, multimodal, with claimed capacity for 100 trillion tokens per day—a figure that, per the AI inference demand index at Tokens Per Day, represents infrastructure at a scale very few organizations can actually field. OpenRouter followed with the same announcement. This wasn't an accidental leak. Someone wanted people talking.

The question worth asking first, before the detective work, is whether this model is actually worth talking about.

What it does, tested

The Better Stack presenter ran the same prompt across seven models: build a full-stack personal finance management dashboard, no stack constraints, just ship a working application. The goal wasn't to trick anyone—it was to see what choices each model makes when left to its own judgment.

Ox Alpha consumed 84,000 tokens and ran for 45 minutes. The output: a functional Next.js application with Drizzle ORM and BetterSQLite. A real database. Working pages. A UI the presenter described as genuinely usable—"if I was making a piece of personal software, I would not complain if it looked like this."

That last part matters more than it sounds. A lot of AI-generated apps are demos: they look like apps, they do not behave like apps. Ox Alpha built something with actual routing, actual persistence, actual architecture decisions that a developer would make.

For free.

Compare that to the paid alternatives in the same test. Fable Medium, at $2.60—the most expensive of the batch—produced a single-page React app with no real database. Refresh the page, lose your data. GPT-5.6 Sol at 33 cents spent four minutes on the task and similarly skipped persistent storage. Gemini 3.7 Flash, run at high effort, built in Next.js but also omitted a real database.

The presenter's read is blunt: "If you're not switching to some of these open models to save yourself some money on that higher thinking, you are definitely missing out."

I'd frame it slightly differently. The closed frontier models aren't necessarily worse—they're operating under different constraints in these tests, mostly time and cost-per-token pressure at medium effort levels. But the value equation has shifted enough that the reflexive assumption—open model equals inferior output—is no longer safe to make.

The benchmark reality check

Before the code tests, there was a benchmark controversy worth unpacking. A tweet circulated claiming Ox Alpha "beats Fable and Soul by quite a bit" on SWE-bench. The presenter tracked this back to a sample of 10 tasks, not a full run. On more complete evaluation, the model sits around 63%—competitive with Grok, DeepSeek, and Gemini, roughly equivalent to Fable and GPT-5.6 Sol at medium effort.

That's still a solid result. But it's not the result being implied by the most excited takes on social media.

This matters because the AI benchmarking ecosystem has a structural problem: partial results travel faster than complete ones, and by the time the correction circulates, the narrative is set. The presenter's willingness to flag this—"anyone telling you this is better than Fable is either clickbaiting or overhyping it, and I'm not here to do that"—is the right instinct. The model is good. It doesn't need to be something it isn't to be worth your attention.

One data point that doesn't fit neatly into either narrative: Ox Alpha averaged roughly 47,000 output tokens on tasks, which is closer to closed model behavior than most open models exhibit. The presenter notes this but doesn't land on an explanation. It's an interesting signal with no clean interpretation yet.

The forensics

This is where it gets genuinely interesting—and where I want to be careful about what's actually established versus what's plausible.

The presenter's method: run Claude across all seven codebases and look for structural similarities and rare markers. The results pointed clearly at GLM 5.3.

The surface-level signal is stack convergence. Of all seven models tested, only Ox Alpha and GLM 5.3 independently chose the identical combination: Next.js, BetterSQLite, and Drizzle ORM. Every other model diverged somewhere. Claude's analysis, as presented in the video, found GLM 5.3 had the highest structural pattern overlap with the stealth model—though the presenter notes this could be inflated by the shared Drizzle usage, and that if you strip out framework overlap, DeepSeek actually scores higher on structural similarity.

What makes the GLM case stick isn't the stack. It's a specific function.

Both Ox Alpha and GLM 5.3 implemented a parseAmountToSense utility function—and the implementations were, according to Claude's analysis, practically identical. Here's why that stops me: this isn't a common utility. It's not something you'd converge on from reading the same documentation. Two models, given the same unconstrained prompt, independently producing the same non-obvious function with the same implementation is not a thing that happens by accident. That's not overlap. That's a fingerprint.

Both models also independently rewrote their README files—something none of the other five models did—and the README structures matched. The presenter links to an external evidence page aggregating additional community analysis, though those specific technical claims remain unverified independently.

The presenter's conclusion: "It is pretty likely from all of this analysis that this is a GLM model or at least it used GLM as a base." The candidates include GLM 5.5 (now multimodal), a post-trained variant of GLM 5.3, or a flash version. And since GLM is open-source, it doesn't have to be ZhipuAI at all—any organization could have built on top of it.

That last point is underappreciated. The stealth launch format, the infrastructure scale, the anonymous branding—these could be a lab testing deployment without reputational exposure, or a company that fine-tuned an open base model and isn't ready to attach its name to it. Neither scenario is sinister. Both are plausible.

Ox Alpha's emergence isn't an isolated event—it echoes a broader pattern of anonymous AI models surfacing with strong capabilities and no lab willing to own them, which suggests this particular playbook is becoming a launch strategy in its own right.

The one thing you should actually know before using it

The launch announcement claimed zero data retention. The presenter flags that this refers to training data: prompts and completions are retained by the provider, but apparently not used to train the model.

"Apparently" is doing real work in that sentence.

For most use cases—experimenting with code generation, building personal tools, testing prompts—this is probably fine. For anything involving sensitive data, proprietary code, or confidential information, "apparently not used for training" from an anonymous provider is not a standard I'd work to. The model is free because someone is paying for it some other way, and until we know who that is and what their data practices actually are, treating it as a zero-privacy-risk tool would be optimistic.

The free window runs through the 27th. Use it for what it's good for. Know what you're trading.