Anonymous AI Model Ox Alpha Outperforms Competitors in Code Tasks
OpenRouter launched the anonymous AI model Ox Alpha on August 20th, offering it for free for one week. The model's performance in code-related tasks has surpassed results from competitors like Claude Fable 5.
The global AI aggregation platform OpenRouter released an anonymous AI model named Ox Alpha on August 20th, making it available for free for one week. The model's performance in reading code and identifying vulnerabilities has drawn significant attention due to its superior results compared to many established models.
In tests conducted by developer Ben Davis, which assessed AI models' ability to read real codebases, identify vulnerabilities, and generate patches, Ox Alpha achieved an average score of 80% across ten tasks. For comparison, Claude Fable 5 scored 65% and GPT-5.6 Sol achieved 52%. Despite the model's unclear origin, its performance has surprised industry observers.
Ox Alpha supports a context window of 1,048,576 tokens and allows for multimodal input, including text, image, and video, as well as handling large outputs of up to 131,072 tokens. The programming agent OpenCode also offers access to the model, claiming it can process up to 100 trillion tokens daily.
Analysis of the model's tokenization and API responses strongly suggests that Ox Alpha may be developed by the Chinese AI company Zhipu. The token count and certain video encoder behaviors align with Zhipu's GLM-5.3 model, providing indications about the model's true origin and developer.