On Aug 20, 2026, an anonymous model called OX Alpha appeared on OpenRouter under stealth/ox-alpha. It has a 1M token context window, accepts text, image, and video input, and costs $0 per million tokens. The operator claims capacity for 100 trillion tokens per day. Community sleuths link it to the GLM family based on tokenizer behavior. Nobody has confirmed anything. The AI industry's first true ghost model.
# The Drop
No announcement. No blog post. No company name. On August 20, a model just appeared on OpenRouter with a slug that said everything: stealth/ox-alpha. Within hours, it was routing traffic from Claude Code users and Hermes Agent sessions.
Model: stealth/ox-alpha Provider: Anonymous third-party Context: 1,000,000 tokens Max Output: 131,072 tokens Modalities: Text, Image, Video (input) Price: $0.00 / $0.00 per million tokens Status: Preview Developer: ???
- ✓ Live on OpenRouter and OpenCode right now
- ✓ Completely free during preview period
- ✓ Multimodal - text, image, and video input
- ✓ Prompts retained by provider, not used for training
- ✗ Who built it - no company, no team, no paper
- ✗ Model architecture - no technical details released
- ✗ Training data - zero transparency
- ✗ Business model - free forever? Or bait?
This is not a beta launch. This is a frontier-class model dropped anonymously onto the busiest AI routing platform in the world. No press release. No waitlist. Just a slug and a million-token context window.
# The Numbers
OX Alpha is not a toy. The performance numbers are real, and the infrastructure behind it is massive. Here is what independent testing and the OpenRouter dashboard reveal.
| Metric | OX Alpha | Notes |
|---|---|---|
| Throughput | 22 tok/s | Competitive with frontier models |
| Latency | 5.74s | Time to first token |
| Cache Hit Rate | ~67% | Prompt caching active |
| Claimed Capacity | ~1.16B tok/s | 100T tokens/day |
| DeepSWE | 80% | Independent, unverified for all repos |
100 trillion tokens per day translates to roughly 1.16 billion tokens per second. For context, that is enough capacity to serve every Claude Code and Hermes Agent user simultaneously. Whoever is behind this has serious infrastructure.
# Who Built This?
Nobody knows. And that is exactly the point. The AI industry has never seen a frontier-class model launch without attribution. Every theory so far is speculation.
Multiple community members report the tokenizer fingerprint matches the GLM family. Video input behavior is also consistent with known GLM capabilities. If true, this could be a Zhipu AI model deployed through a proxy. But Zhipu has not claimed it, and the operator has not confirmed anything.
A well-funded stealth startup using OpenRouter as a live stress test before a formal launch. The "stealth" slug supports this reading. Free access generates real-world usage data at scale without marketing spend.
Free frontier model plus "prompts retained by provider" equals massive prompt collection at zero marginal cost. Agentic coding tools route complex, multi-step prompts. That is extremely valuable training signal. The operator's privacy policy says prompts are not used for training - but the operator is anonymous.
Every other frontier model has a company behind it. A legal entity. A terms of service you can enforce. OX Alpha has an anonymous operator and a privacy policy nobody can hold anyone to. The model performs well. The question is whether "performs well" is enough when you cannot verify who receives your prompts.
# Why This Changes Everything
OX Alpha is not just a model. It is a category event. The AI industry assumed frontier models required public companies, billion-dollar funding rounds, and marketing campaigns. That assumption just died.
Training a frontier model no longer requires revealing who you are. Routing platforms like OpenRouter enable zero-attribution launches. Expect more ghost models.
A 1M-context multimodal model at $0 per token undercuts every commercial API. Subsidized or not, it forces competitors to justify their pricing.
The EU AI Act requires frontier model providers to identify themselves. OX Alpha is a live test case for whether that framework actually works when the provider is a ghost.
OX Alpha's strengths - coding, frontend, sustained agentic work - are exactly what Claude Code and Cursor users need. This is a targeted play for the highest-value segment.
# Via OpenRouter API 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": "Build a React dashboard with auth"} ] }' # Cost: $0.00 # Context: 1,000,000 tokens # Developer: Unknown
OX Alpha is the AI industry's first ghost model. A frontier-class system with no company, no paper, no attribution - just a slug on OpenRouter and enough infrastructure to serve 100 trillion tokens per day. The performance is real. The 80% DeepSWE score is promising. The $0 price tag is aggressive. But the fundamental question remains: should you route your code through a model when you cannot verify who receives it? The answer depends on your threat model. The fact that the question even exists is the story.
Claude's Invisible Ink: Anthropic Watermarks All Output
Everything Claude writes now carries an invisible fingerprint. Not just in the EU - everywhere, for everyone. The watermark travels with copy-paste. But can it survive the real world?
Enjoyed this?
New episodes Mon, Wed, Sat.