Mario

I think we may have been looking at Ox Alpha the wrong way. Giving away absurd amounts of frontier inference for free never really looked like charity. Then I checked OpenRouter’s Stealth Model EULA. It literally says stealth models can be offered free for the purpose of collecting user content for training, evaluation and improvement. So suddenly the whole strategy makes a lot more sense: → anonymous model → insane free usage → hundreds of thousands of developers stress-test it → real-world coding tasks at massive scale → feedback + eval signals → model gets refined before the reveal → enormous organic marketing at the same time And interestingly, several of us have noticed Ox producing dramatically better results on similar prompts than it did on day one. Now, an important caveat: I cannot prove Ox is continuously training on those prompts in real time. In fact, Ox’s specific OpenRouter listing says retained data is not used for training, while OpenCode advertises its route as ZDR/no-training. The published policies are messy enough that I would not make the stronger claim yet. But as a massive real-world evaluation + launch flywheel? Yeah. This may have been one of the smartest stealth-model rollouts I’ve seen.