Ox Alpha is a reasoning-focused AI model released on August 20, 2026, designed specifically for long-horizon software engineering and agentic workflows [S1, S4]. Available through OpenRouter under the identifier stealth/ox-alpha, the model is currently offered at zero cost for both input and output tokens during its preview period [S1, S4, S7].
Unlike established models from named labs, Ox Alpha is a “stealth” release, meaning the developer and training details remain officially undisclosed [S6, S8]. While it is accessible via an OpenAI-compatible API, it is operated by an anonymous third-party provider rather than OpenRouter itself [S6, S8].
Technical Capabilities and Agentic Support
The defining feature of Ox Alpha is its 1,048,576-token context window [S1, S2, S4]. This capacity allows developers to load entire large codebases, documentation, and task histories into a single prompt, reducing the need for the model to infer context from isolated snippets [S1, S3]. The model supports a maximum output of 131,072 tokens, providing room for structured plans and detailed technical solutions [S1, S4].
Beyond text, Ox Alpha is multimodal, accepting images and video as input [S1, S4, S6]. This enables practical debugging workflows where an agent can analyze a browser screenshot, a UI mockup, or a screen recording of a bug alongside the relevant source code [6].
For agentic work, the model supports tool calling, tool choice, and structured JSON output [S1, S4, S8]. These features allow it to function within autonomous loops that alternate between planning, gathering evidence via tools, and deciding the next action [1].
Privacy Risks and Data Retention
Because Ox Alpha is a stealth model, privacy terms vary significantly depending on the access route. OpenRouter states that while prompts and completions are retained by the provider, they are not used for training [S2, S4]. However, other sources note that the general stealth-model agreement may grant broader rights, and some listings contain contradictory language regarding training rights [S2, S3].
In contrast, the OpenCode Zen route lists the model as x-preview-f-free and claims zero retention and no training [2].
Given these inconsistencies, developers should treat Ox Alpha as a temporary evaluation tool rather than a default destination for private or regulated source code [2]. It is recommended to test the model on redacted repositories or fixed tasks before any team-wide rollout [2].
Forensic Identity and Performance
While the provider remains anonymous, independent serving-layer forensics published on August 22, 2026, point toward Zhipu AI’s Z.ai infrastructure [S3, S5]. Researchers identified a Java stack trace naming internal Zhipu API classes and a specific error-code dialect (code 1214) that matches Z.ai-hosted GLM models but differs from other hosts of the same weights [S3, S5].
Additional evidence includes a 30-of-30 match on tokenizer probes to GLM-5.3 and video-encoder signatures that align with GLM-5V-Turbo [S3, S5].
Preliminary community benchmarks suggest high performance in coding tasks. In a 10-task trial on DeepSWE, Ox Alpha reportedly achieved an 80% pass rate, surpassing some frontier models in that specific limited set [3]. Other benchmarks indicate strong accuracy in mathematics (97%) and coding (95%), though instruction following (77%) is a relative area for improvement [7].
Practical Integration Guide
Ox Alpha can be integrated into existing workflows using an OpenAI-compatible base URL (https://openrouter.ai/api/v1) and the stealth/ox-alpha model slug [S4, S8].
To maintain security during evaluation, follow these steps:
- Store API keys in server-side environment variables rather than browser code [S1, S8].
- Use the
reasoning_effortparameter (low, high, or max) to balance latency and quality [8]. - Validate all generated code with standard build and security review processes, as context capacity does not replace ownership of the final change [1].
- Design agents with explicit checkpoints and require human confirmation for irreversible operations [1].
Developers can currently test these capabilities through the OpenRouter API or integrated agent harnesses like Claude Code and Hermes Agent [4].
Sources
- Ox Alpha on OpenRouter: Free 1M Stealth Model (Aug 2026) - explainx.ai
- Ox Alpha AI Model: Anonymous 1M-Context Coding Agent
- Ox Alpha - Free AI Model for Coding & Agentic Work
- Ox Alpha Free AI Model: Setup, Privacy and Cost | Wavect
- Ox Alpha - AI Model Details & Benchmarks
- Coding Model Ox Alpha Retains Every Prompt: You Cannot Name Company …
- Ox Alpha: What We Know About the Mystery AI Model
- Ox Alpha AI — 1M Context, API & OpenRouter Guide
