Mysterious AI model ox alpha hits openrouter with strong performance in deepswe benchmark

The OpenRouter platform has launched an anonymous AI model called Ox Alpha, which is currently available for free for one week. The model has demonstrated strong performance in the DeepSWE benchmark, which measures an AI's ability to read real codebases, identify vulnerabilities, and generate effective patches. Ox Alpha scored approximately 80% in the test, significantly outperforming Claude Fable 5, which achieved 65%, and GPT-5.6 Sol, which scored 52%. Developer Ben Davis expressed surprise at the results, noting the model's unknown origin. Ox Alpha is equipped with a context window of 1,048,576 tokens and supports multimodal input, including text, images, and video. The maximum output per response is 131,072 tokens. The model is available for free during its preview period. OpenCode, a terminal programming intelligence agent, also offers the model through its free tier, claiming it can process up to 100 trillion tokens daily. Researchers conducted a tokenizer fingerprint analysis on 25 prompts and found that Ox Alpha's token count closely matches GLM-5.3, with only a constant 75-token wrapper difference. Additionally, the video encoder behavior and API response characteristics of Ox Alpha align closely with the architecture of Zhipu AI. These findings have led to speculation that the model may be developed by Zhipu AI, though no official confirmation has been made.