Mysterious Model ox-alpha Surpasses GPT-5.6 in Coding Benchmarks, Technical Fingerprints Point to Zhipu AI's Unreleased New Model
Nashnova编辑部
An anonymous model called ox-alpha hit 80% Pass@1 on a coding benchmark, beating Claude and GPT-5.6; an independent researcher matched its technical fingerprints to Zhipu's unreleased next-gen flagship, with a late-August reveal window now in focus.
Where did this model come from?
On August 20, an anonymous model tagged stealth/ox-alpha quietly appeared on the model-routing platform OpenRouter, with no entity claiming ownership.
Independent researcher Ben Davis ran systematic tests and said on X he is "99% certain" ox-alpha belongs to Zhipu's (智谱) GLM-5.x family.
This means → if the attribution holds, Zhipu is stress-testing its next-generation multimodal flagship under cover before an official launch.
What makes the case for Zhipu?
Davis presented evidence across four dimensions: video encoder, tokenizer, audio interface, and output style — with the video encoder considered the strongest signal.
In four controlled video tests, ox-alpha's video-token consumption exactly matched GLM-5V-Turbo's — both showed frame-rate-independent sampling, a scaling ratio of roughly 147 tokens per second, and per-frame resolution scaling. Rival candidates MiMo v2.5, Qwen 3.8 Max, and GLM-4.6V all diverged.
The tokenizer lined up too: across 25 prompts, ox-alpha's token counts were identical to GLM-5.3, with only a fixed +75-token hidden wrapper offset. In plain terms = the two models likely share the same vocabulary table.
ox-alpha also refused audio input, matching GLM-5V's behavior — while the leading alternative candidate, MiMo v2.5, accepts audio, further weakening its case.
How strong is it at coding?
On the DeepSWE coding benchmark, ox-alpha passed 8 of 10 deterministic tasks — a Pass@1 rate of 80%.
For comparison: Claude Fable 5 scored 65%, GLM-5.3 and Grok 4.6 both hit 62%, and GPT-5.6-sol managed just 52%.
On one task (meriyah-explicit-resource-declarations) where every competitor had gone 0/4, ox-alpha passed on the first attempt. This means → its capability sits clearly above GLM-5.3, looking more like a next-generation checkpoint than a minor variant.
How large is the model, and who is running it?
ox-alpha's decoding speed is within roughly 6% of GLM-5V-Turbo, which has 744B total parameters and 40B active parameters.
Davis infers ox-alpha likely uses a similarly scaled MoE architecture — MoE, or mixture-of-experts, meaning only a fraction of the model's parameters fire on each query rather than all at once.
In plain terms = if active parameters are indeed around 40B, the operator's claimed capacity of 100 trillion tokens per day starts to make sense on both cost and compute grounds.
When will we know for sure?
Zhipu has a track record of stealth-channel model testing — Pony Alpha was eventually confirmed as linked to GLM-5.
ox-alpha is currently in a free-access window that may last until August 27. Previous stealth models were formally claimed by their Chinese AI labs once free testing ended.
This reflects an emerging industry playbook: release anonymously, harvest real-user feedback and benchmark data, then announce on your own terms — the days around August 27 are shaping up as the key confirmation window.
Content is for reference only, not financial advice.