Anonymous AI model going wild: Could Ox Alpha be hiding Zhipu’s next-gen GLM? An AI model with no company, no launch event, and not even a public identity suddenly steps up to challenge GPT-5.6, Claude, and Grok in side-by-side tests. Its name? Ox Alpha. Since landing on OpenRouter on August 20, this “ghost model” has taken the developer community by storm. 1 million tokens of context, support for image and video input, 128K output, high-intensity inference enabled by default, and most importantly: It’s free. But what’s really shaking up the community isn’t the specs—it’s how it performs on real-world coding tasks. Someone tested it on DeepSWE’s real software engineering tasks: Ox Alpha: 80% Fable 5 Max: 65% GLM-5.3: 62% Grok 4.6 xhigh: 62% GPT-5.6 Sol Max: 52% 10 tasks, completed 8. Sure, the sample size is small, so we can’t outright declare it “the model to beat,” but the results are enough to make developers start dissecting it like crazy. And here’s the big question: Who—or what—is Ox Alpha? The community hasn’t found an official website, but they’ve uncovered a ton of technical fingerprints. In tokenizer tests, Ox Alpha is highly consistent with GLM-5.3; Its token consumption pattern for video input is similar to the GLM-5V series; And its API parameters and inference behavior show a lot of similarities too. The leading theory right now? It might just be Zhipu’s next-gen multimodal GLM model currently in testing: http://Z.ai.