Ox Alpha, an unattributed AI model released on August 20 through OpenRouter, OpenCode, Cline and Nous Research’s Portal+, processed 11.6 trillion tokens during its first 72 hours on OpenRouter, about 2.6 times the previous debut record. It was free during a preview that ended on August 24, although an OpenCode launch note had described the period as lasting about a week. The model offered a 1,048,576-token context window, roughly 131,000-token maximum outputs, text, image and video input, tool and function calling, zero data retention and generous rate limits, making it attractive to coding agents. Developer Ben Davis recorded an 80% score on a 10-task DeepSWE sample, above Claude Fable 5 at 65% and GPT-5.6 Sol at 52%, but two full 113-task runs placed Ox Alpha near 63%, roughly matching GPT-5.6 Sol and showing the small sample’s volatility. No organization has claimed the model. Independent fingerprinting matched its tokenizer and video-token behavior to Zhipu AI’s GLM-5.3 family, suggesting an unreleased multimodal system sometimes called GLM-5.3 Flash, but Zhipu has not confirmed the connection.