subsidized versus API pricing sure, but do we have concrete proof of how much it is subsidized versus the breakeven inference costs of those tokens?