The anonymous “stealth” AI model OX Alpha appeared on OpenRouter and OpenCode on August 20, 2026, immediately capturing the AI community’s attention. With a massive 1M-token context window, free access, and an unknown creator, it has become one of the most discussed AI releases of the year.
What Is OX Alpha Stealth Model?
OX Alpha is an anonymous AI model that appeared on the OpenRouter marketplace on August 20, 2026, listed under the “stealth” provider category . It is described as a reasoning model designed for coding, sustained agentic work, and production workloads .
Key confirmed specifications :
- Context window: 1,048,576 tokens (approximately 1 million tokens)
- Maximum output: 131,072 tokens
- Modalities: Text, image, and video input (text-only output)
- Reasoning: Enabled by default; cannot be disabled
- Tool calling: Supported
- Structured output: JSON format supported (strict JSON schema not yet supported)
- Price during preview: Free
- Release date: August 20, 2026
- Platform availability: OpenRouter (stealth/ox-alpha), OpenCode (x-preview-f-free), CommandLyne.AI
The model is explicitly described by OpenRouter as being “developed and operated by a third-party model provider who has chosen to remain anonymous during this preview” . OpenRouter routes requests to the provider but is not the developer, owner, or provider of the model .
Why Is OX Alpha Being Called a Stealth AI Model?
The “stealth” designation stems from the model’s anonymous release. Unlike traditional AI launches accompanied by company announcements, blog posts, and technical papers, OX Alpha appeared without any formal introduction .
Key factors contributing to the “stealth” label:
- Undisclosed developer: No company or organization has claimed ownership
- No launch announcement: No press release, blog post, or technical report accompanied the release
- Limited public information: Model parameters, training data, architecture, and benchmark results remain confidential
- Sudden platform appearance: The model was simply uploaded to OpenRouter and OpenCode
- Anonymous provider label: OpenRouter lists it only as “stealth”
OpenRouter co-founder Alex Atallah characterized this as “recently the first Stealth release” on social media .
OX Alpha’s 1M-Token Context Window Explained
The model’s advertised 1,048,576-token context window is among its most notable features. In practical terms, this means OX Alpha can process text equivalent to multiple full-length novels in a single session.
How the large context window benefits users:
- Large codebases: Developers can feed entire code repositories into the model for comprehensive analysis and debugging
- Long documents: Processing lengthy reports, research papers, or legal documents in one go
- Multi-step tasks: Maintaining coherence across extended workflows without losing context
- Agentic workflows: Supporting sustained agentic work where the model continuously reads files, calls tools, tracks modifications, and executes tests
- Complex reasoning: Handling extended chains of thought for sophisticated problem-solving
The 1M context window places OX Alpha in the same tier as leading models from companies like Google and OpenAI .
OX Alpha AI Model Features
| Feature | Details |
|---|---|
| Model type | Reasoning model |
| Context window | 1,048,576 tokens (1M+) |
| Reasoning | Enabled by default; cannot be disabled |
| Coding | Primary focus; optimized for software engineering tasks |
| Multimodal | Text, image, and video input support |
| Agentic tasks | Designed for sustained agentic work |
| API availability | OpenRouter, OpenCode, CommandLyne.AI |
| Platform availability | OpenRouter, OpenCode |
| Pricing | Free during preview |
| Release/public availability | August 20, 2026 |
How Good Is OX Alpha?
Independent testing has provided mixed signals about the model’s actual performance.
Early DeepSWE benchmark results :
Developer Ben Davis conducted a 10-task DeepSWE benchmark test, which measures an AI model’s ability to read real codebases, identify bugs, and generate valid fixes. The early results:
| Model | DeepSWE Score (10 tasks) |
|---|---|
| OX Alpha (stealth) | 80% |
| Claude Fable 5 (max) | 65% |
| GLM-5.3 (max) | 62% |
| Grok 4.6 (xhigh) | 62% |
| GPT-5.6 Sol (max) | 52% |
A subsequent independent developer tested OX Alpha on a complete 113-task DeepSWE benchmark, which took over 20 hours to complete. The model solved 66 tasks, achieving a completion rate of 58.4% (95% confidence interval: 49.2% to 67.1%). This placed OX Alpha close to Claude Opus 4.8 at 59% .
Important context:
- The 80% early result came from a small sample size and should not be treated as definitive
- About 80% of tasks achieved over 90% pass rates; approximately 10% of failures resulted from tool-calling format errors rather than logic errors
- No official benchmark results from organizations like Artificial Analysis have been published
- The small sample results may not represent comprehensive capabilities
- The model’s identity, architecture, and training data remain unknown, making benchmark interpretation complex
OX Alpha vs GPT-5.6, GLM and Other AI Models
Note: The DeepSWE comparisons are based on a small sample and different test configurations. Reliable, comprehensive comparisons are not yet available due to limited independent benchmark data .
Who Made OX Alpha?
CONFIRMED: OX Alpha’s creator or underlying organization has not been officially confirmed .
OpenRouter explicitly states: “Ox Alpha is developed and operated by a third-party provider who has chosen to remain anonymous during this preview” .
What the AI Community Is Speculating
SPECULATION: The strongest community theory points to Chinese AI company Zhipu AI (Z.ai), specifically an unreleased multimodal version of its GLM-5.3 model .
Reported evidence cited by researchers :
- Tokenizer matching: Researchers found that OX Alpha’s tokenization method for CJK characters, emojis, and Cyrillic letters matched Hugging Face’s published GLM tokenizer
- Token count consistency: A 25-prompt comparison showed OX Alpha’s token counts differed from GLM-5.3 by a constant 75 tokens, suggesting system prompt packaging
- Error message: The 1210 error returned when attempting to disable reasoning mode matches Zhipu’s error patterns
- Video encoder behavior: OX Alpha’s multimodal behavior reportedly resembles Zhipu’s GLM-5V-Turbo
Other community theories mentioned in reports include :
- DeepSeek-V4-Flash variant
- SpaceXAI’s unreleased Grok model
- Microsoft MAI-2 (supported by some citing the use of OpenAI’s cl100k_base tokenizer)
- Xiaomi MiMo V3
- Tencent’s new model
Important: All of these are community speculations. Zhipu AI has not confirmed involvement. The model’s true developer remains officially undisclosed .
Speculation Rule: All information about the developer outside of OpenRouter’s official statement remains unconfirmed and should be treated as community speculation unless verified by official sources.
Why the “OX Alpha” Name and Sudden Free Access?
The name “Ox Alpha” (often referred to in Chinese communities as “牛來,” a phonetic play on “Ox is Coming”) emerged alongside an extremely generous free access model.
Usage statistics from the first three days :
- OpenRouter: ~336 billion prompt tokens, 4 billion response tokens
- OpenCode: ~1.8 trillion tokens, 57,000 users, 865,000 conversation sessions
- Top applications: Hermes Agent, Claude Code, and ZCode accounted for most traffic
- Anthropic’s Claude Code alone used OX Alpha for 314 billion tokens over three days
The free period has been announced as approximately one week by OpenCode (varying between “future week” and “future 6 days” depending on announcement), estimated to end around August 27, 2026 .
The estimated server costs for this free usage have been described in industry terms as substantial, with some reports suggesting the cost would exceed billions of Korean won if sold for a fee .
How to Use OX Alpha
Verified access methods:
- OpenRouter – Model ID:
stealth/ox-alpha - OpenCode – Available through the platform’s free tier
- CommandLyne.AI – Free access
Important usage notes:
- The model is free during the preview period
- OpenCode has announced free access for approximately one week
- The model provider states that “input content will not be used for training”
- However, terms of service allow developers to retain input data
- Given the developer’s unknown identity, there is no way to independently verify compliance
Security Warning: As this is a stealth model with limited information publicly available about its origins and data handling, exercise caution regarding what information you share and your data’s confidentiality . Avoid inputting business secrets, unpublished code, or personal data .
Is OX Alpha Free?
Current pricing: Free during the preview period .
Important caveats:
- The free access is temporary
- OpenCode announced free access for “approximately one week” from August 20
- Estimated free period end date: around August 27, 2026
- No official pricing information has been published for after the free period
- OpenRouter currently displays the price as zero with no specific end date listed
Who Can Benefit From OX Alpha?
- Developers and programmers: Coding assistance, code analysis, debugging
- AI researchers: Testing a new model’s capabilities, especially for long-context tasks
- Software engineering teams: Long-horizon software engineering projects
- Agentic AI developers: Building and testing AI agents for complex workflows
- Students: Free access to a high-capability model for learning and experimentation
- Businesses: Exploring AI capabilities for their workflows
OX Alpha Limitations
- Unknown developer identity: Users cannot verify who created the model or their data handling practices
- Unknown architecture: Model parameters, training data, and infrastructure remain undisclosed
- Uncertain pricing future: No official pricing for after the free period
- Limited benchmarking: Comprehensive independent benchmark data is not yet available
- Potential privacy concerns: Data handling cannot be independently verified
- No rigorous JSON schema support: Only basic JSON structured output is supported
- Reasoning cannot be disabled: Attempting to turn off reasoning returns error 1210
- Temporary free access: Users should prepare for potential paid tiers
OX Alpha Latest Update
Date: August 23, 2026
Latest verified developments:
- OX Alpha remains available on OpenRouter and OpenCode as of August 23, 2026
- The model continues to generate significant community discussion and usage
- Independent testing has produced mixed benchmark results, with early 80% scores now contextualized by a more rigorous 58.4% DeepSWE performance
- OpenCode’s free period is expected to end around August 27, 2026
- No official announcement from the model’s creator has been made
- Community speculation continues to point toward various potential developers, most notably Zhipu AI
The situation remains fluid, with new testing, speculation, and potential developments emerging daily.
FAQ
1. What is the OX Alpha Stealth Model?
OX Alpha is an anonymous AI model that appeared on OpenRouter on August 20, 2026. It features a 1M-token context window, multimodal input support, and is designed for coding and agentic tasks. Its creator remains undisclosed .
2. Who created OX Alpha?
The creator has not been officially confirmed. OpenRouter states it is operated by a third-party provider who chose to remain anonymous. Community speculation points to Chinese company Zhipu AI (GLM-5.3) but this has not been confirmed .
3. Is OX Alpha an open-source AI model?
No, OX Alpha is not open-source. It is an anonymous proprietary model available through platforms like OpenRouter and OpenCode .
4. Does OX Alpha have a 1M-token context window?
Yes, the model has a confirmed 1,048,576-token context window .
5. Is OX Alpha available on OpenRouter?
Yes, OX Alpha is available on OpenRouter under the model ID stealth/ox-alpha .
6. Can developers use OX Alpha for coding?
Yes, OX Alpha is specifically described as a reasoning model designed for coding, sustained agentic work, and software engineering tasks .
7. Is OX Alpha free?
Currently, yes. The model is free during the preview period, with OpenCode offering free access for approximately one week from August 20, 2026 .
8. How can I access OX Alpha?
OX Alpha can be accessed via OpenRouter (stealth/ox-alpha), OpenCode (x-preview-f-free), and CommandLyne.AI .
9. Is OX Alpha better than GPT-5.6?
Limited benchmark data is available. A small 10-task DeepSWE test showed OX Alpha scoring 80% vs GPT-5.6 Sol’s 52%, but these results are from a small sample. Comprehensive comparisons are not yet available .
10. Is OX Alpha related to GLM?
This is community speculation. Tokenizer and behavior matching have led researchers to point to Zhipu AI’s GLM series, but this has not been officially confirmed .
11. Why is OX Alpha called a stealth model?
Because it was released anonymously without a formal launch announcement, developer disclosure, or technical paper. It appeared on platforms under the “stealth” provider label .
12. Is OX Alpha safe to use?
The model provider states input will not be used for training, but terms allow data retention. Since the developer’s identity is unknown, data handling cannot be independently verified. Exercise caution with sensitive information .