Ox Alpha AI: Free 1M Context Reasoning Model (2026)

On August 20, 2026, an unannounced frontier reasoning model named Ox Alpha (model identifier: stealth/ox-alpha) appeared on OpenRouter. The model is completely free ($0.00 per 1M input and output tokens) during its initial public preview and features a massive 1,048,576-token context window (1M) alongside native video, image, and text input support.

According to OpenRouter’s official listing, Ox Alpha is “a reasoning model designed for coding, sustained agentic work, and production workloads. It is suited for long-horizon software engineering, complex reasoning, and workflows that combine text with visual context.”

Within hours of deployment, autonomous developer toolchains (including Claude Code, Cline, and Zed) processed hundreds of millions of tokens through the stealth endpoint. Here is the technical breakdown of Ox Alpha: core specifications, early benchmark evaluations, privacy policies, and who is likely behind the model.

---

1. Ox Alpha Technical Specifications

Ox Alpha represents the first stealth reasoning model on OpenRouter to offer native video processing alongside a 1M token context buffer:

Specification	Parameter	Details
Model Identifier	stealth/ox-alpha	Available via OpenRouter API and OpenCode
Pricing	Free ($0 / 1M tokens)	$0 input / $0 output during preview window
Context Window	1,048,576 tokens (1M)	Full 1M token input memory buffer
Max Output Tokens	131,072 tokens	Supports extensive single-pass codebases
Input Modalities	Text, Images, Native Video	First stealth model with native video input
Output Modality	Text only	Generates code, structured JSON, and reasoning traces
Reasoning Mode	Mandatory	Configurable effort: low, high, max (default: max)
P50 Latency / Speed	~3.15s - 3.98s	~29 - 30 tokens/second throughput
Function Calling	Supported	Compatible with OpenAI tool calling schema

---

2. Early Benchmark Performance and Coding Evals

While official evaluations from Artificial Analysis are pending, early community benchmarks on software engineering benchmarks place Ox Alpha among top frontier models:

DeepSWE 10-Task Subset (Independent Community Evals)

On a standardized 10-task repository-level debugging subset of DeepSWE:

• Ox Alpha: 80.0% Pass@1
• Fable 5: 65.0% Pass@1
• GLM-5.3: 62.0% Pass@1
• Grok 4.6: 62.0% Pass@1
• GPT-5.6 Sol: 52.0% Pass@1

Kingbench Multi-Task Reasoning

• GLM-5.3: 91.25%
• Ox Alpha: 87.50% (2nd place)
• Claude Opus 4.8: 84.10%
• Qwen 3.8 Max: 83.50%

Early testing indicates strong instruction adherence on long-horizon refactoring tasks, automated multi-turn error correction, and robust token conservation during complex planning loops.

---

3. Who Built Ox Alpha? Stealth Model Fingerprinting

Ox Alpha follows a well-established pattern of anonymous “stealth drops” on OpenRouter originating from major frontier labs:

• Pony Alpha was later confirmed as Zhipu AI’s GLM-5.
• Hunter Alpha resolved to Xiaomi’s MiMo-V2-Pro.
• Elephant Alpha was revealed as Ant Group’s Ling-2.6.
• Owl Alpha launched as Meituan’s LongCat-2.0.

Primary Evidence: Zhipu AI (GLM-5.3V / GLM-5.5)

Tokenizer analysis and API behavior tests strongly point to Zhipu AI (Z.ai):

1. Tokenizer Alignment: Byte-pair encoding tests show a 50/50 token alignment with the GLM-5.3 tokenizer.
2. Video Token Budgets: The internal frame-sampling and video token allocation match the exact architecture of GLM-5V-Turbo.
3. Reasoning Schema: System prompts and reasoning effort toggles mirror Zhipu’s internal testing configurations.

A secondary candidate is the Xiaomi MiMo research team, though Zhipu remains the highest-probability developer.

---

4. Privacy, Data Retention, and Production Suitability

Before routing proprietary codebases through stealth/ox-alpha, engineering leads should review OpenRouter’s stealth model data policy:

• Data Retention: Prompts and completions are retained by the anonymous provider for evaluation and diagnostic purposes.
• Model Training: The provider policy specifies that inputs are not used for model training.
• Preview Availability: The free tier is scheduled to run for approximately one week (through late August 2026), after which standard commercial API pricing will apply.

Recommendation: Ox Alpha is ideal for non-sensitive code refactoring, side-project scaffolding, and agent evaluation. Teams with strict SOC2 or IP indemnification requirements should wait for official lab de-anonymization and enterprise SLA terms.

---

5. How to Use Ox Alpha (Setup Guide)

Ox Alpha is accessible immediately through any OpenAI-compatible client, IDE extension, or terminal assistant:

1. Direct API Integration (cURL / Python / Node)

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "stealth/ox-alpha",
    "messages": [
      {"role": "user", "content": "Refactor this database connection pool in Go"}
    ]
  }'

2. Integration in Claude Code, Cline, and Zed

1. Obtain an API key from openrouter.ai.
2. Set the base URL in your configuration to https://openrouter.ai/api/v1.
3. Set the target model ID to stealth/ox-alpha.

---

Frequently Asked Questions

Is Ox Alpha completely free?

Yes. During its public preview window on OpenRouter, Ox Alpha has zero input token fees and zero output token fees ($0/1M tokens).

What is the context window of Ox Alpha?

Ox Alpha features a 1,048,576 token (1M) context window and supports up to 131,072 output tokens per completion.

Can Ox Alpha process video files?

Yes. Ox Alpha is the first stealth model on OpenRouter to natively accept video, high-resolution image sequences, and text inputs.

---

Last updated: August 21, 2026. Sourced from OpenRouter model cards and community benchmark evaluations.

Primary Sources:

• OpenRouter Ox Alpha Model Page: openrouter.ai/stealth/ox-alpha
• OpenRouter Stealth Model Terms: openrouter.ai/terms