Ox Alpha is presented as an anonymous, preview-stage AI model aimed at coding, complex reasoning, and long-running agent workflows. It surfaced quietly on OpenRouter with no company attribution, parameter count, or technical report, and was listed free during its preview window. Technical fingerprinting reported by third-party researchers points toward the GLM family lineage: the model matched 44 out of 44 discriminating tokenizer strings against GLM-5.2 and GLM-5.3, exposed reasoning level controls similar to GLM-5, and its video-token behavior aligned closely with GLM-5V-Turbo. The official developer remains undisclosed in published write-ups, but those tokenizer, infrastructure, and multimodal signals have been cited as strong circumstantial evidence of a Z.ai origin.
In benchmark testing focused on software engineering, Ox Alpha posted an 80% completion rate on a ten-task DeepSWE sample, ahead of Claude Fable 5 (65%) and GPT-5.6 Sol (52%) in that single independent run, and a 58.4% score across the fuller 113-task DeepSWE suite, placing it near Claude Opus 4.8 (59%). It also completed 28% of LiveCodeBench v6 problems and 17.9% of SlopCodeBench strict checkpoints, performing well on isolated tasks but showing some quality drift during sustained repository maintenance. Practical use cases span large codebase analysis, multi-file debugging, long document review, tool-driven workflows, and multimodal inputs that combine text, images, and video, with the service exposed through a standard OpenAI-compatible Chat Completions endpoint.