Mysterious 'Ox Alpha' Model Sparks Identity Debate: GLM or Gemini?

From the moment it appeared, developers have been trying to figure out who trained it. The leading candidates are Z.ai (the company behind GLM) and Google. Here's how the evidence stacks up for the "Z.ai" theory. A widely circulated test showed that for the same mixed text containing Chinese, English, code, and emojis, GLM-5.3 and GLM-5.2 both calculated 68 tokens, while Ox Alpha produced 143. The extra 75 tokens could be explained by an external hidden system prompt. Another test with 25 different prompts showed a similar fixed offset. More striking is the video token fingerprint. Researchers created four controlled test videos and compared Ox Alpha with several multimodal models. Ox Alpha and GLM-5V-Turbo showed highly similar token budgets, including sampling strategy unaffected by frame rate changes, approximately 147 tokens per second growth with duration, and identical token scaling for resolution changes. The extra token budgets matched item for item. Even the API layer left clues: an error message like "[1210] The temperature parameter is illegal" resembles Z.ai's service style. Some users even claimed to have captured a residual "trained by Z.ai" string in the output, though that is only from community screenshots and secondhand accounts, and is considered a lower-level piece of evidence.

The Google theory is largely driven by product profile and recent social media activity by DeepMind employees. Ox Alpha's 1M context, native image/video input, long-horizon agent focus, and strong tool-calling match Google's Gemini strengths. Google has yet to release Gemini 3.5 Pro, which was originally promised "next month" during I/O and has been delayed. DeepMind researcher Evan Otero posted vaguely about "Gemini" and said "if Ox Alpha is... a friend we've met along the way," which many interpreted as a hint. Techmeme also noted that Gemini team members have recently been discussing Gemini and next-generation models more frequently.

As of now, there is no official confirmation from either company. The public evidence suggests that external behavior alone is increasingly insufficient to determine the developer, especially with distillation, model merging, and third-party deployment becoming common.

Ox Alpha itself has been put through rigorous testing. It has accumulated over 250B tokens of usage, with about 8,500 unique users and more than 110,000 completed sessions, according to OpenCode. Stripe CEO Patrick Collison called it "very impressive" after testing. In a community test of 10 real software engineering tasks, Ox Alpha completed 8 (80%), compared to Fable 5 at 65%, GLM-5.3 at 62%, and GPT-5.6 Sol at 52%. However, such tests are small-sample and not official benchmarks. Sceptics include Abacus.AI CEO Bindu Reddy, who said Ox Alpha performed below expectations, and Wharton professor Ethan Mollick, who called it only moderately impressive. Some developers with private benchmarks also noted weak performance at low reasoning intensity, and visual tasks appeared considerably weaker than Gemini models.

A common joke suggests that Cursor might have trained a model on open-source GLM and wrapped it in a "cow head." The answer is likely to come soon, as according to the pattern of previous OpenRouter alpha models, the anonymous period usually ends with someone claiming the "cow."