Article: Ox-Alpha, an unnamed model tagged only as “stealth/ox-alpha,” appeared on OpenRouter on 20 August 2026. It offers a 1,048,576-token context window, video input, and a free preview that ends on 27 August, inviting developers to test a capability that can swallow entire codebases or multiple novels in a single request.

OpenRouter has become a hub for experimental models that surface without a corporate banner, often labeled “stealth.” The platform’s low-friction publishing pipeline lets researchers and labs release prototypes under a generic tag, bypassing the usual branding and support structures. Ox-Alpha follows that pattern, joining earlier anonymous releases such as Pony Alpha and Hunter Alpha, which later traced back to Chinese research labs.

Why the token window matters

Most commercial large-language models (LLMs) cap their context at around 8 k tokens. A million-token window expands what can be fed to the model by an order of magnitude. In practice, developers can:

  • paste an entire software repository and ask for cross-file refactoring,
  • feed a full manuscript and request a chapter-by-chapter summary,
  • combine text and video frames to reason about UI design or visual bugs.

The model also accepts video as a first-class input, a feature still rare among LLMs that usually rely on static images or separate video-specific networks.

Early performance signals

A researcher who ran the DeepSWE benchmark—a suite that evaluates coding ability across several languages—recorded an 80 % score. The benchmark does not directly compare Ox-Alpha to established models such as GPT-5.6 or Claude Opus 5, but the result signals strong competence in programming tasks.

In a separate agentic workflow test, the model executed 69 tool calls and produced only a single error. The test measured reliability when orchestrating external APIs, a key metric for autonomous assistants that need to chain actions without human oversight.

Who might be behind it?

No organization has claimed ownership, but technical forensics point toward a possible connection with Zhipu AI’s unreleased GLM-5.x series. The clues include:

  • video-encoding patterns that match those used in known GLM models,
  • tokenizer behavior aligning with the GLM-5.3 token set,
  • a distinctive style of emoji usage seen in prior GLM releases.

If the attribution is correct, Ox-Alpha would represent a leap from Zhipu’s internal prototypes to a publicly accessible service, echoing earlier “stealth” releases that eventually disclosed Chinese origins.

What developers can gain

  • Large-scale code analysis – run static analysis or refactoring across thousands of files without chopping the input.
  • Visual-context coding – combine video frames with code generation for UI automation or video-editing scripts.
  • Long-context workflows – build agents that retain memory over extended interactions, useful for research assistants or narrative generation.

Risks and unknowns

  • No documentation – the model lacks official guides, making integration a trial-and-error process.
  • No support or service-level agreements – any downtime or unexpected behavior falls on the user.
  • Pricing – the model is free during the preview period, which ends on 27 August.

Takeaway

Ox-Alpha shows that powerful, long-context AI capabilities can appear overnight, sidestepping traditional branding and support structures. For developers, the model offers a tempting shortcut to tasks that previously required custom pipelines, but the lack of provenance, documentation, and guaranteed service demands a cautious approach. The coming week will reveal whether the mystery behind Ox-Alpha resolves into a sustainable tool or remains a fleeting experiment.