Who Made Ox Alpha? Every Clue, From Tokenizer Fingerprints to Pony Alpha
No lab has claimed ox-alpha, the free stealth coding model on OpenRouter. We rank every attribution clue — the 44/44 tokenizer match to GLM-5.3, the compression analysis, the self-identification leak, and the theories that don't hold up.
Short answer: nobody has claimed it, and the strongest evidence points to a GLM-5.3-generation model from Z.ai (Zhipu AI) — but that's high-confidence inference, not confirmation. Here is every public clue, ranked by how much it actually tells you.
The timeline that frames everything
- Aug 14, 2026 — Z.ai launches GLM-5.3, text-only.
- Aug 20, 2026 — an unattributed model appears on OpenRouter as
stealth/ox-alpha: free, 1M context, multimodal input (text, images, and video), built for coding and agentic work. It soon surfaces on OpenCode, Cline, and Nous Research's portal as well. - Aug 20–25 — independent attribution investigations begin; the tokenizer evidence lands.
- Since then: silence. Zhipu AI has made no statement confirming or denying.
Connect the Claude or Codex you already pay for — the rest runs on workers that cost a fraction.
Download meshcode →The clues, ranked by strength
1. Tokenizer fingerprinting — the strongest evidence. Community tooling built for exactly this situation (the open-source modelprint approach) compares how a model tokenizes text, which is nearly impossible to fake and tracks the training pipeline. One public investigation (LuD1161/ox-alpha-identification-public) reports a 44/44 match on tokenizer differentials against GLM-5.3, concluding ox-alpha is a Z.ai GLM-5-generation model with high confidence. Every alternative candidate scored noticeably worse on the same tests.
2. Compression-based stylistic attribution. A separate dejan.ai analysis used compression-based attribution over writing samples; GLM-5.3 was the top match (7/14 samples). Different method, same direction.
3. The self-identification leak. Reportedly, when ox-alpha's own system prompt was fed back to it, it identified itself as "GLM, made by Z.ai." On its own this proves little — models parrot their prompts all the time — but as consistent with the fingerprinting it adds weight.
4. The Pony Alpha precedent. Z.ai has run this exact playbook before: GLM-5 previewed anonymously on OpenRouter under the name "Pony Alpha." Anonymous stealth previews are an established part of this lab's launch marketing, which makes the hypothesis self-consistent.
5. Positioning. Coding, long-horizon agentic work, production workloads, a free preview with massive capacity — the shape matches where Zhipu has been pushing the GLM line all year.
The wrinkle: video input
Here's the part that keeps the story open. GLM-5.3 launched text-only on Aug 14; ox-alpha accepts video input. The leading explanation: ox-alpha isn't GLM-5.3 but an unreleased multimodal upgrade of it — the community has taken to calling it "GLM-5.3 Flash" — or a newer checkpoint entirely. That would also explain the anonymous preview: stress-test capacity and real agent workloads before putting a name on it.
The theories that don't hold up
Microsoft's unreleased MAI family, Xiaomi's MiMo line, DeepSeek, Alibaba's Qwen, and Google Gemini were all floated in the first days. On tokenizer similarity tests, most of these candidates scored poorly. They remain speculation; the GLM family is the only one with positive evidence behind it.
Why attribution actually matters
Two practical reasons this isn't just gossip:
- Compliance. Zhipu AI was added to the U.S. Commerce Department's Entity List in January 2025. If ox-alpha is Zhipu, some enterprises — and especially some U.S. government-adjacent teams — have procurement rules that change the answer to "can we use this?" That question is worth answering before a free preview becomes load-bearing.
- Data flow. The provider retains your prompts and completions (published terms say they're not used for training). An anonymous provider with retention rights deserves the same caution you'd give any unknown vendor — fine for experiments and fan-out work, think first before pasting production secrets.
What would settle it
Any of: a statement from Zhipu AI; a GLM-5.3 multimodal release whose fingerprints match; or the preview ending with a named reveal, Pony-Alpha-style. Until one of those happens, "probably Z.ai, not confirmed" is the honest position.
Want the performance side of the story? We broke out the benchmark numbers — including why the viral "80% on DeepSWE" claim doesn't survive full runs — in Ox Alpha benchmarks, explained. And if you just want to use the thing while it's free: run ox-alpha in MeshCode in about a minute.
Attribution status as of August 26, 2026. The developer has chosen to remain anonymous during the preview; no theory above is confirmed. MeshCode is not affiliated with Z.ai, Zhipu AI, or OpenRouter.