Who Is Behind Ox Alpha? OpenRouter's Stealth Model, Explained

A frontier model with no name attached

On August 20, 2026, a model called Ox Alpha appeared on OpenRouter under the identifier stealth/ox-alpha. It arrived with a 1,048,576-token context window, video input, mandatory reasoning — and no lab attached. The provider field says only "Stealth." The tokenizer field says only "Other." The Hugging Face field is empty. Nobody has claimed it.

It is also, for now, free — input and output both priced at $0 during a preview period that multiple reports say ends around August 27. By the weekend, OpenRouter's usage stats showed the model had already processed trillions of tokens. Stripe CEO Patrick Collison called it "very impressive." That's not a chatbot demo; that's production traffic.

One naming note before you search anything: it's Ox Alpha, as in the animal. If you've seen "0x Alpha" written with a zero, that's a typo — and it drags in results about the unrelated 0x Protocol DeFi project.

Note: attribution for this model is contested, not settled. Treat everything below as evidence, not verdict — this piece separates what the platform metadata actually says from what researchers have inferred.

What the API metadata actually says

Rather than quote anyone's summary, you can pull the raw catalogue entry yourself:

curl -s https://openrouter.ai/api/v1/models \
  | jq '.data[] | select(.id=="stealth/ox-alpha")'

Three of those fields deserve emphasis. hugging_face_id: null means there are no open weights — the "Ox-Alpha-GGUF" repository that appeared on Hugging Face is a community placeholder with no model files in it. knowledge_cutoff: null means the "cutoff around November 2025" figure circulating in write-ups is somebody's inference from probing the model, not a published fact. And the pricing line reads $0 only "while the preview runs" — OpenRouter's terms let the provider pull the model "at any time... with or without notice."

The fingerprint evidence: it looks like GLM-5.3

You can hide a model's name. It's considerably harder to hide its behavior — and models leave fingerprints.

Serving profile. OpenRouter exposes more metadata than most people read. When researchers compared Ox Alpha against all 422 models in the catalogue, only three entries combine mandatory reasoning with the max/high/low effort ladder and a default of max: stealth/ox-alpha, z-ai/glm-5.3, and the glm-latest alias pointing at it. That's a match on nine of ten serving fields.

Tokenizer behavior. One developer ran side-by-side tokenization tests — English, German, Chinese, code, emoji. Ox Alpha produced exactly the same tokenization pattern as Z.ai's GLM-5.3 every time, except it consistently contained 75 additional tokens. A fixed difference like that is consistent with the same tokenizer wrapped around a hidden system prompt. For the full picture of that model family, see our GLM-5.3 review.

Deterministic outputs. With randomness dialed down and identical prompts, researchers reported striking similarities in formatting, phrasing, and output structure between Ox Alpha and the GLM family.

None of this proves Z.ai owns Ox Alpha. But at some point you stop finding fingerprints and start finding the driver's license in the getaway car.

The competing theory: Microsoft

A separate line of analysis — built on cl100k-family tokenizer comparisons — points at an unreleased Microsoft model (MAI or a Phi successor). The two camps reached different conclusions by different routes, and as of August 24, no lab has claimed the model. What everyone agrees on: someone is paying real money to let the world hammer on a frontier-class model for free, anonymously.

The benchmark that got retracted

Here's the part most coverage hasn't caught up with. The number that made Ox Alpha go viral — 80% on DeepSWE, ahead of GPT-5.6 Sol (52%), GLM-5.3 (62%), and Claude Fable 5 (65%) — came from a 10-task subset. The same author re-ran the full 113-task benchmark and got roughly 63%. The 80% figure was withdrawn by the person who produced it.

Ten tasks is directional evidence, not a verdict. A model that's genuinely competitive-but-not-dominant is still a very interesting free preview — it's just not the "MYSTERY CHINESE AI DESTROYS OPENAI" headline.

The catch: your prompts are training data

This is the part that should give every builder pause. OpenRouter's Stealth Program terms grant the provider a "perpetual" license to pass your content to the unnamed operator so it can "train, evaluate, and improve" the model. Three separate policy statements about this contradict each other. Combine that with is_moderated: false and an unvetted operator, and you have a model you should think twice before sending proprietary code to — even at $0.

For the framework thinking about which model should handle which task (and what each tier costs you), see our AI agent cost control guide.

Should you use it?

Practically, in the remaining preview days:

1. Do experiment. A free frontier-class model with a 1M-token context and native tool calling is a rare test bench. Long-document summarization, multi-file code analysis, video-input tasks — try the things your current model can't hold in context.
2. Don't send anything sensitive. Assume every prompt becomes someone's training data, because the terms literally say it might.
3. Don't build on it. It can disappear "at any time, with or without notice" — and the $0 pricing almost certainly ends this week. Anything you wire into production needs a fallback model behind an abstraction layer.
4. Watch the reveal. Whoever it is, the launch playbook — anonymous frontier model, free tokens, community forensics doing the marketing — is now proven. Expect more of these.

If Ox Alpha really is a GLM-5.3 variant in a trench coat, the interesting question isn't "who" but "why": why give away a frontier model anonymously? The answer is probably competitive benchmarking at scale — 6 trillion tokens of free, real-world stress testing before an official launch.

Bottom line

Ox Alpha is a genuinely capable model wrapped in a genuinely unusual launch: no lab, no weights, no published cutoff, no moderation, and a data license you should read twice. Use the free window to learn what a 1M-token context actually unlocks — then read the fingerprint evidence and decide for yourself who's behind the curtain.

Related: GLM-5.3 review · AI agent cost control · DeepSeek Harness overview · Hugging Face guide · AI Agent Guide

Next Steps

Compare the fingerprint case against GLM-5.3, then decide whether a stealth preview belongs anywhere near production traffic.

how do AI agents work — return to the complete AI agent architecture guide.