On August 20, an anonymous AI model named Ox Alpha appeared on the Portal platforms of OpenRouter, OpenCode, Cline, and Nous Research, and was made available for free use. Due to the absence of any corporate attribution, this "stealth model" quickly garnered attention within the developer community.
The initial surge in interest was driven by a set of rapidly spreading code benchmark results. After testing on 10 tasks from DeepSWE, developer Ben Davis reported an 80% first-pass success rate, surpassing Claude Fable 5’s 65% and GPT-5.6 Sol’s 52%. This result was widely shared, leading many to believe that Ox Alpha had clearly outperformed existing leading models.
Full testing returned to approximately 63%
DeepSWE is a benchmark for code agents that measures whether models can solve real GitHub issues on the first attempt. It consists of 113 tasks spanning 91 code repositories.
Based on two independent, complete benchmark runs mentioned in the text, Ox Alpha's scores fell at approximately 63%, roughly on par with GPT-5.6 Sol, rather than consistently outperforming it. In other words, the previously widely circulated claim of an “80% lead” was largely based on small-sample results with significant variability.
As of August 22, Ox Alpha has not yet appeared on major model rankings such as Artificial Analysis or LMSys Arena, meaning that publicly available evaluation data about it remains limited.
Free access drives engagement
Beyond performance discussions, the open conditions stated by the platform have also increased attention. OpenCode stated that its routing can support a processing volume of 100 trillion tokens per day; Nous Research claimed that its Portal can handle up to 1 quadrillion tokens per day.
- Open for free for approximately one week
- Supports a 1 million context window
- Supports multimodal input and zero data retention
Stripe CEO Patrick Collison also posted on X, calling the model "very impressive." This brief comment further amplified the discussion around Ox Alpha.
Fingerprint analysis points to Zhipu AI
No institution has yet publicly claimed ownership of Ox Alpha. External parties have conducted multiple rounds of comparisons regarding its origin.
The report mentions that several common candidates have been partially ruled out. For instance, Xiaomi MiMo v2.5 supports audio input, while Ox Alpha explicitly rejects audio; DeepSeek has never previously offered video capabilities and typically releases open-source weights rather than conducting anonymous previews first; Google and Qwen also differ from Ox Alpha in their tokenizers and video encoding methods.
The remaining clues point more strongly to Zhipu AI. Independent researchers' fingerprint tests revealed that Ox Alpha's tokenizer is identical to GLM-5.3 in standardized tests, its video token consumption perfectly matches GLM-5V-Turbo across four video samples, and it also exhibits similar error patterns and the same behavior of rejecting audio input as the GLM series.
However, this judgment still has one point of uncertainty. When GLM-5.3 was released on August 14, it was still a text-only model, while Ox Alpha launched on August 20 with full video capabilities. Based on this time gap, if the above fingerprint analysis holds, Ox Alpha is more likely an unpublished multimodal upgrade of the GLM-5.3 series rather than a direct copy of an existing model.
The so-called "stealth model" typically refers to a model developed by an organization that remains anonymous, initially deployed through routing platforms like OpenRouter to collect real-world usage data before deciding whether to officially release it. The report notes that this practice has become relatively common among some Chinese AI labs. According to OpenCode's announcement, the free open period for Ox Alpha is expected to last until approximately August 27, and whether its true identity will be disclosed at that time remains one of the market's key points of interest.