Ox Alpha: A Stealth Reasoning Model for Coding and Agentic Workflows

Ox Alpha is a multimodal reasoning model for production engineering

Ox Alpha is a reasoning model designed for coding, sustained agentic work, and production workloads. It is specifically optimized for long-horizon software engineering, complex reasoning tasks, and workflows that require the integration of text with visual context. Released on August 20, 2026, the model is currently available as a "stealth" model on OpenRouter, meaning it is developed and operated by an anonymous third-party provider.

Technical Specifications and Capabilities

Ox Alpha provides high-capacity context and multimodal input support to facilitate complex software development and agentic tasks.

Context and Modalities

  • Context Window: 1,048,576 tokens.
  • Maximum Output: 131,072 tokens.
  • Input Modalities: Supports text, images, and video.
  • Output Modalities: Returns text.

Feature Support

  • Tool Calling: Supports tools and tool_choice for function calling.
  • Structured Outputs: Supports response_format for JSON output (without JSON-schema enforcement).
  • Reasoning: Supports reasoning-enabled requests, allowing users to access the model's internal thinking process via the reasoning_details array in the API response.

Performance and Pricing

Ox Alpha is currently offered for free through OpenRouter, with no charges for prompt or completion tokens.

Latency and Throughput

Based on OpenRouter's P50 metrics:

  • Throughput: 33 tokens per second (tps).
  • Latency: 3.76 seconds.
  • Availability: 98.47% over a three-day period.
  • Cache Hit Rate: Average of 72.97%.

Production Usage

High-volume traffic to Ox Alpha is driven by agentic coding tools and autonomous agents, including:

  • Claude Code: 27.7B tokens
  • Hermes Agent: 21.7B tokens
  • Oh-My-Pi: 19.9B tokens
  • DeepSeek Harness: 17.2B tokens
  • ZCode: 13.4B tokens

Community Analysis and Speculation

Because the model's provider remains anonymous, the developer community has used stylometry and behavioral analysis to speculate on its origin.

Probable Origin: GLM Series

Multiple users suggest the model is a variant of the GLM (General Language Model) series, citing specific reasoning patterns and stylistic markers.

"Based on its indecisive and far-too-lengthy thinking traces when given complex instructions... as well as a rudimentary stylometry... this is almost certainly a GLM model."

Other theories suggest it could be a GLM-5.3 Flash or a vision-enabled variant of GLM-5.3, potentially intended to compete with DeepSeek's flash models.

Behavioral Observations

  • Knowledge Cutoff: Testing suggests a knowledge cutoff around mid-2025.
  • Content Filtering: Users report the model refuses to answer questions regarding Tiananmen Square but is more permissive regarding technical instructions for electronic warfare that other frontier models typically block.
  • Coding Performance: While some users find it capable for general coding, others report failures in specific tasks, such as implementing CSS color-mix() functions, where the model allegedly replaced dynamic logic with hardcoded hexadecimal values.
  • Creative Tasks: Some users report the model outperforms other high-end models (like "K3") on creative and "looser" tasks, though visual reasoning is noted as a weaker point.

Data Privacy and Terms

Ox Alpha is governed by OpenRouter's Stealth Model Terms. While the provider states that prompts and completions are retained but not used for training, the anonymity of the provider has led to community caution regarding the submission of proprietary or confidential data.

Sources

Related

  • Dispatch
    Z.AI Ox Alpha: New GLM-Series Model with Open Weights

    Z.AI (Zhipu) has confirmed that Ox Alpha is a new iteration of its GLM series and has committed to releasing the model's weights to compete with DeepSeek.

  • Dispatch
    Qwen3.7-Max agent model release

    Qwen released Qwen3.7-Max, a new agent-focused foundation model that excels at coding, office automation, and ultra-long-horizon autonomous tasks, now available via Alibaba Cloud Model Studio.

  • Project
    zai-org/GLM-5

    GLM‑5 is an open‑source series of very large LLMs (320 B‑744 B total parameters) optimized for multimodal input, 1 M‑token contexts, and long‑horizon agentic tasks. It introduces hybrid sparse/linear attention, IndexShare, and a custom RL engine (slime) to achieve frontier‑level performance on coding and planning benchmarks while offering cheap FP8 inference. Models are downloadable from Hugging Face/ModelScope and can be served with SGLang, vLLM, KTransformers, etc., with optional ‘reasoning_effort’ control.

  • Dispatch
    OpenAI o3 and o4-mini release notes / what's new

    OpenAI has released o3 and o4-mini, reasoning models that integrate multimodal capabilities and agentic tool use to solve complex problems across coding, math, and science.

  • Dispatch
    AI & Frontier Tech Roundup: The Shift Toward Agentic Workflows and Local Sovereignty

    The AI landscape is pivoting from standalone chat models toward autonomous agentic workflows and high-performance local inference to bypass cloud limits and costs.