Mysterious Ox Alpha AI Model Debuts With One-Million-Token Context Window

A powerful but anonymous artificial intelligence model has appeared on OpenRouter, attracting immediate attention from developers for its enormous context window, multimodal capabilities, and focus on autonomous coding.

The new model is called Ox Alpha and is listed under the identifier stealth/ox-alpha. It first appeared on OpenRouter on August 20, 2026, before becoming available through OpenCode as a limited-time free model.

Unlike conventional AI product launches, Ox Alpha arrived without an identified developer, an official research paper, or a public announcement from the laboratory that created it. The organisation operating the model remains undisclosed.

What Is Ox Alpha?

Ox Alpha AI is a multimodal reasoning model designed primarily for software engineering, complex reasoning and long-running agentic workloads.

According to its public specifications, the model supports:

• a 1,048,576-token context window;
• text, image and video inputs;
• text output;
• tool calling;
• structured responses;
• up to 131,072 output tokens;
• extended coding and software-engineering tasks.

The model is currently accessible through the official Ox Alpha page on OpenRouter.

Its one-million-token context window means that Ox Alpha can theoretically process extremely large volumes of information during a single task. This could include an extensive software repository, its documentation, development history, application logs and visual references.

Large context capacity does not guarantee that a model will understand every piece of information accurately. It can, however, reduce the need for developers to divide a complex project into numerous disconnected prompts.

Ox Alpha Targets Autonomous Coding

Ox Alpha appears to have been developed for more than generating isolated pieces of code.

Its official positioning emphasises long-horizon software engineering and sustained agentic work. These are assignments in which an AI system must plan several steps, access tools, inspect files, modify code, run tests and correct its own mistakes.

Adding image and video input could allow an AI coding agent to analyse screenshots, interfaces, application behaviour, and visual errors alongside the underlying source code.

In practice, an agent powered by a model such as Ox Alpha could:

1. inspect a large software repository;
2. identify the files related to a reported problem;
3. plan and implement a correction;
4. run automated tests;
5. examine the result;
6. revise the implementation when necessary;
7. document the completed changes.

This is the type of complete digital workflow that could reduce the human labour required for routine development, testing, maintenance and technical documentation.

A detailed analysis of the model, its access options and potential applications is available in the complete Ox Alpha AI guide published by Replace Humans.

Who Created Ox Alpha?

The company behind Ox Alpha has not been officially confirmed.

Its anonymous launch has generated several theories within the developer community. Some technical observers have associated the model with Z.ai’s GLM model family, while others have suggested a connection to Xiaomi’s MiMo team.

These theories are based on characteristics such as tokenizer behaviour, output patterns, multimodal functionality and serving performance. None amounts to an official disclosure.

OpenRouter should not be confused with the model’s developer. The platform provides access to AI models from multiple organisations, while the “Stealth” designation allows an unidentified provider to test a model without publicly revealing its name.

Anonymous previews can help AI companies collect feedback, measure real-world performance and test infrastructure before launching a model under its final commercial identity.

Ox Alpha Is Currently Free

Ox Alpha is currently listed at no cost for input or output tokens during its preview period.

OpenCode also lists Ox Alpha Free among its available models for a limited time. Permanent pricing has not been announced, and the current free access should not be interpreted as a commitment that the model will remain free.

Offering an advanced model without usage charges can rapidly attract developers and generate an enormous variety of real-world tasks. The provider can then study performance across coding environments, agent frameworks and large-context workloads before deciding how the final product will be released.

Businesses testing Ox Alpha should therefore avoid building permanent cost estimates around the preview price.

Can Ox Alpha Compete With GPT-5.6 and Claude?

Ox Alpha’s specifications place it in the same conversation as leading models used for coding and autonomous agents. Its one-million-token context window, multimodal input and support for tool use make it technically relevant to workloads commonly assigned to GPT-5.6, Claude and Gemini.

However, there is not yet enough independently verified evidence to declare Ox Alpha superior to those models.

A reliable coding comparison must measure more than a model’s ability to solve individual programming problems. It should evaluate whether the model can:

• complete entire assignments;
• follow repository-specific instructions;
• use tools correctly;
• avoid introducing regressions;
• recover from failed actions;
• remain consistent during long tasks;
• produce working code across repeated tests.

Ox Alpha may prove highly competitive, but early community experiments should not be treated as definitive performance benchmarks.

Questions Remain About Privacy and Production Use

Ox Alpha’s current information indicates that it does not use prompts and completions to train the model. Businesses should still examine the applicable retention policies before uploading confidential information.

A commitment not to train on customer data is not necessarily the same as zero data retention.

Until the operator is identified and permanent commercial terms are available, companies should avoid submitting proprietary repositories, credentials, customer records or regulated information. Initial evaluations can be performed using public codebases, synthetic tasks or controlled internal data.

The model’s anonymous origin also leaves several questions unanswered:

• Who operates the infrastructure?
• Will Ox Alpha receive an official name?
• How much will it cost after the preview?
• Will its model weights be released?
• What enterprise support will be available?
• What data-processing terms will apply to commercial users?

Why the Ox Alpha Launch Matters

The importance of Ox Alpha extends beyond its developer.

The model reflects a broader transition in artificial intelligence. AI systems are no longer being designed only to answer questions or produce short pieces of content. They are being optimised to understand larger working environments and continue executing tasks with less human supervision.

A coding model capable of examining an entire repository, interpreting visual results, using development tools and correcting its own work can absorb responsibilities currently distributed among developers, testers, code reviewers and technical writers.

Ox Alpha has not yet proved that it can replace those workers reliably. Its design demonstrates where the market is moving.

The decisive competition between AI laboratories is increasingly about which model can complete the longest and most valuable sequence of work without requiring human intervention.

Ox Alpha has entered that competition anonymously—but it is unlikely to remain unnoticed.