Ox Alpha, an unattributed AI model focused on coding and sustained agentic work, launched on August 20 through OpenRouter, OpenCode, Cline and Nous Research’s portal+. It initially went viral after developer Ben Davis recorded an 80% score on a 10-task DeepSWE sample, ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%. Two separate full 113-task runs, however, placed Ox Alpha near 63%, roughly level with GPT-5.6 Sol and underscoring the volatility of the smaller sample. The model is free during an initial window of about a week, supports text, image and video input, offers roughly 1 million tokens of context and tool and function calling, and has zero data retention and generous rate limits. Its JSON output is not schema-enforced. OpenCode said it could handle 100 trillion tokens per day, while Nous Research claimed capacity for 1 quadrillion. Stripe CEO Patrick Collison called the model very impressive after his company agreed on August 19 to acquire OpenRouter. Although developers floated Google, Xiaomi, DeepSeek, Microsoft’s MAI, Alibaba’s Qwen and other candidates, independent fingerprinting linked Ox Alpha’s tokenizer to Zhipu AI’s GLM-5.3 and its video-token behavior to GLM-5V-Turbo. Because GLM-5.3 launched as text-only on August 14, analysts suspect Ox Alpha may be an unreleased multimodal upgrade, sometimes referred to as GLM-5.3 Flash. The model had no listing on Artificial Analysis or LMSys Arena as of August 22, and its likely free-access cutoff was around August 27.