I have not see any reporting or evidence at all that Anthropic or OpenAI is able to make money on inference yet.
> Turns out there was a lot of low-hanging fruit in terms of inference optimization that hadn't been plucked yet.
That does not mean the frontier labs are pricing their APIs to cover their costs yet.
It can both be true that it has gotten cheaper for them to provide inference and that they still are subsidizing inference costs.
In fact, I'd argue that's way more likely given that has been precisely the goto strategy for highly-competitive startups for awhile now. Price low to pump adoption and dominate the market, worry about raising prices for financial sustainability later, burn through investor money until then.
What no one outside of these frontier labs knows right now is how big the gap is between current pricing and eventual pricing.
