A mysterious AI model called Ox Alpha has appeared on OpenRouter without revealing the company behind it, offering developers free access to a reasoning model built for coding, long-running AI agents and production workloads.
The model surfaced on August 20 under OpenRouter’s ‘Stealth’ label. Its provider has chosen to remain anonymous, with OpenRouter saying it only routes requests to the third-party provider.
Ox Alpha comes with over 1 million token context window, a maximum output of 131,072 tokens, and support for text, image, and video inputs. It also supports tool calling and structured outputs, putting it squarely in the growing category of models built for agentic software development rather than simple chatbot interactions.
The model is currently available for free, with OpenCode promoting it as a limited-time preview. The provider has reportedly claimed capacity for as much as 100 trillion tokens per day, while OpenCode's listing describes the model as free for a week.
Coding Performance Draws Attention
Ox Alpha quickly gained attention after early developer testing suggested strong coding performance.
In a test conducted by independent developer and researcher Ben Davis, Ox Alpha scored 80% after completing eight of 10 DeepSWE software-engineering tasks. On the same comparison, Claude Fable 5 scored 65%, GLM-5.3 62%, Grok 4.6 62% and GPT-5.6 Sol 52%.
The model’s ability to work across long software-engineering tasks has been enough to attract significant developer interest.
Who Built Ox Alpha?
That is the biggest question surrounding the launch.
The strongest theory so far points to Z.ai, the Chinese AI company behind the GLM family. Researchers have compared Ox Alpha’s API behaviour, tokenisation, and error responses with publicly available GLM models and found several similarities.
Ananay Arora, Founder of OpenHack, said on social media that Ox Alpha is “100% a GLM model by Zhipu AI” and suggested that the model could be an upcoming GLM-6.
Based on early benchmark results, Arora speculated in a post on X that Ox Alpha appeared to be “almost Mythos class” and was outperforming other frontier models in software engineering and cybersecurity benchmarks.
Besides that, one independent fingerprinting effort found Ox Alpha’s tokeniser matched the GLM-5 generation across 44 test strings. Other researchers found similarities in API behaviour and reasoning controls.
The connection to Z.ai remains unconfirmed, with one notable clue standing out. While GLM-5.3 is a text-only model, Ox Alpha can process images and videos, fuelling speculation that it could be an unreleased multimodal model from the GLM family rather than a rebranded GLM-5.3.
Anonymous model releases are not entirely new. Z.ai previously used the name Pony Alpha for an anonymous preview before the release of GLM-5, giving developers a chance to test the model without the usual branding and expectations.
Ox Alpha has also seen substantial usage since its release. OpenCode’s tracking page currently shows more than 22 trillion tokens processed, 282,000 unique users, and nearly 6.8 million completed sessions.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.