what model is ox alpha on openrouter
Ox Alpha is almost certainly Z.ai’s GLM-5.x family wearing a mask. No one has officially claimed the anonymous stealth model that appeared free on OpenRouter around August 20–22, 2026 — but independent fingerprinting keeps landing on the same answer: tokenizer counts matching GLM-5.3, video-token behavior matching GLM-5V-Turbo, and GLM-style quirks. It’s a stealth preview, not a surprise new frontier lab.
what is ox alpha on openrouter?
A “stealth” model — released under a codename with no company attached — listed as stealth/ox-alpha. It landed the week of August 20, 2026 with headline specs: a ~1-million-token context window (1,048,576), up to 131,072 tokens of output, and multimodal text/image/video input. It’s free during a short promotional window that ends around August 27. Labs do this to A/B a model against the public without the branding and expectations of an official launch — and to soak up real-world traffic for evaluation.
which model is ox alpha, really?
The interesting part isn’t the specs, it’s that people reverse-engineered its identity from behavior alone. Fingerprinting a black-box model means measuring things the operator can’t easily fake — how it tokenizes, how it eats video, what it refuses, how it writes. Every independent signal so far converges on Z.ai / GLM:
| signal | what testers measured | points to |
|---|---|---|
| tokenizer | raw token counts match GLM-5.3 across 25 prompts, offset by a fixed ~75-token wrapper | Z.ai GLM-5.3 |
| video encoder | ~147 tokens per second of video, frame-rate-independent sampling | GLM-5V-Turbo |
| audio | rejects audio input the same way GLM-5V does | GLM family (rules out audio models like MiMo v2.5) |
| output style | ~1.3 emojis per 1,000 characters | GLM/Qwen (Claude, GPT-5.6, Grok sit near zero) |
| error strings | GLM/Qwen-style “dirty token” errors | GLM/Qwen lineage |
None of these is a confession. But a near-perfect tokenizer match is hard to fake by accident — the tokenizer is baked into how a model was trained — and when the video budget, the audio refusal, and even the emoji rate all point the same way, coincidence gets expensive. One researcher put his confidence at 99% GLM-5.x. The honest read: almost certainly GLM-family, very likely served on Z.ai-related infrastructure, officially unclaimed.
could it be something other than glm?
Yes, and a good post says so. The strongest counter comes from Sonar researcher Joe Tyler, who points out that “dirty token” errors “are intrinsic to the way modern LLMs are trained and used for inference, not something specific to one provider.” In other words, some of the softer fingerprints are shared across many Chinese models, not unique to GLM. There’s also a live alternative candidate: Xiaomi’s MiMo team, which has floated stealth models on OpenRouter before under names like “Hunter Alpha.” So the truly rigorous version is: the tokenizer and video evidence make GLM the overwhelming favorite, but “confirmed” needs Z.ai to say the word — and it hasn’t.
is ox alpha free, and what’s the catch?
It’s free in the only way stealth models are ever free: you’re the eval set. Zero dollars per token during preview, but every prompt you send routes to an operator with no name on the door and no terms you can hold anyone to. That’s the actual story behind the “retains every prompt” headlines — it’s not a scandal, it’s just what a black-box preview is. Ship it toy problems and coding tasks; don’t paste your codebase’s crown jewels, customer data, or anything you’d mind a mystery lab in another jurisdiction keeping.
should you use it?
For throwaway coding and kicking the tires on a 1M-context multimodal model at $0, sure — it benchmarks well (around 80% Pass@1 on DeepSWE). Just use it for what it is: a free preview of what’s very likely GLM-5.x, logged by someone who won’t sign their name. Enjoy the free tokens, keep your secrets to yourself, and don’t call it a mysterious new frontier lab — the tokenizer already told on it.