I would guess the real value of the position Openrouter has in the ecosystem is that they can intercept all data and sell training data from many people, training data that is about as directly relevant for training as it can get. They would not need to log all your data for this, "sampling" is enough, and might in fact be even better sampled.

Note: this would include both the request, the response, and competitive responses. This would be very high quality training data.

Now at the current point they're sporting and make this data sampling option opt-out, and indeed they do. By default, they collect your data.

Model companies will pay a lot for this, especially now, more than customers will ever give to openrouter. Think about this: this is a constant stream of data of exactly what people and companies do with AI models, for all models, updating LIVE.

Second, they can provide "competitive intelligence" as well, telling models live how well they're doing and what exactly their strong and weak points are, and in what distribution of token requests this results.

There may even be financial companies that pay for this data.