A stealth AI model called Ox Alpha surfaced on OpenRouter on August 20, offering a 1,048,576-token context window, 131,072-token max output, tool calling, structured outputs, and text, image, and video input, priced at zero dollars per million tokens. No developer has claimed it. The listing runs under the provider name "stealth," and OpenRouter is explicit that it only routes traffic, it doesn't own or build the model.
That anonymity is the entire marketing strategy, and it's working. Stripe CEO Patrick Collison called it "very impressive" after testing it. A subset test of 10 DeepSWE coding tasks put it at an 80% pass rate, ahead of Claude Fable 5 and GPT-5.6 Sol, though the full 113-task suite brought that down to a more ordinary 63%, roughly tied with GPT-5.6. That gap between the viral number and the full number is worth sitting with before anyone declares a new frontier leader.
Ox Alpha = A now-familiar playbook, done at unusual scale
Unnamed "stealth" previews aren't new. OpenRouter has run them before, and two past stealth models turned out to be Xiaomi's MiMo systems once revealed. What's different here is the scale of the offer: OpenCode reported the provider has capacity for 100 trillion tokens a day, free, for about a week. That's not a small lab quietly testing a model on a few hundred users. That's an operation confident enough in its compute to give away a frontier-scale model at internet scale, purely to harvest usage data and buzz before a real launch.
The guessing game over who's behind it has produced more noise than signal. A DeepMind researcher's cryptic post fueled Gemini rumors, but tokenizer behavior, error signatures, and system artifacts point more consistently toward Zhipu AI's GLM-5.3, per independent technical forensics. None of it is confirmed, and OpenRouter isn't required to say.
What this means if you're building on these tools
The free-and-anonymous model era is a live audition, not a stable option. Treat any stealth preview, this one included, as a time-boxed experiment: useful for testing what a frontier model can do at your workload before committing, not something to route production traffic through long-term. The terms of use, data retention, and even the model's existence can change without notice, because by design, nobody's accountable for it yet.
If your team is trying to figure out which of these tools are worth building workflows around versus which are just noise, that's exactly the kind of call our growth strategy work is built to make with you, and our AI marketing services team can help you put the reliable ones to actual use.
Writing Team
.png)