Ox Alpha, a free AI reasoning model designed for coding, sustained agentic tasks, and production workloads, was anonymously released on OpenRouter on August 20, 2026 [1, 2]. The model supports very long context windows of up to 1,048,576 tokens and can process text, images, and video inputs, though not audio [2]. It claims an exceptionally high usage capacity, handling up to 100 trillion tokens per day, free until at least August 24 or August 27 depending on the source [1, 2].
The model is described as a "stealth model," developed and operated by an unnamed third-party provider who declined to disclose their identity during its preview period [3, 1, 4, 2]. OpenRouter serves only as the routing platform and is not the developer or operator of Ox Alpha [2]. According to the provider, user inputs and outputs are retained but not used for training, although OpenRouter’s broader terms permit data use for evaluation and improvement [2].
Early benchmark tests showed Ox Alpha achieving over 80% on the DeepSWE software engineering benchmark, outperforming Anthropic's Claude Fable 5 at 65% and OpenAI’s GPT-5.6-sol at 52%, although these results were based on small samples and showed variability [2]. The model’s output token limit per request is 131,072 tokens [2].
Speculation about Ox Alpha’s developer has focused on the Chinese AI company Z.ai (formerly Zhipu AI), due to tokenizer similarities with the GLM-5.3 model released by Z.ai on August 14, 2026, just six days before Ox Alpha appeared [3, 1, 4, 2]. AI analyst Andrew Curran noted on August 23 that "this morning people seem less sure of anything" regarding firm attribution [3, 1]. Other candidates proposed include Microsoft’s unreleased MAI models and Xiaomi’s MiMo team, both linked by shared technical fingerprints or previous anonymous previews [3, 1, 4, 2]. However, some experts caution that matching fingerprints may indicate common infrastructure rather than a definitive developer identity, as developer unclecode stated: "指紋匹配證明的是共用基礎設施，而不是身分" (fingerprint matches prove shared infrastructure, not identity) [2].
On August 23, prominent technology figures including Stripe CEO Patrick Collison praised Ox Alpha’s performance, calling it "very impressive" [3].
Free access through OpenRouter was initially set to end on August 24, 2026, but OpenCode announced a one-week extension of free usage until August 27, 2026 [2]. This extension allows wider testing and use of the model during the final days of its free preview.
Ox Alpha’s anonymous release with its advanced capabilities and enormous token capacity has attracted significant attention as users evaluate its potential and origins over the coming days.
