i think the concerns about ai subscription subsidies are overblown. yes, as far as i can tell, the $200 / month claude max plan supports something like $2000 / month in api token costs if ur pushing it really hard as a coding agent. but the $20 plan, not so much, the rate limits there are quite low, and very likely it remains profitable due to many users not even coming close to those limits. generally, inference at the api token price is clearly profitable, and likely *very* profitable. many businesses will and are happily paying the thousands a month for peak coding agent usage. over time the subsidy might drop, but unlike past VC subsidies the product has a clear path to getting cheaper and better over time. uber now is just the same thing but more expensive, but sonnet 4.6 has pretty similar coding perf to opus 4.5 and faster speed and lower cost, and this trajectory is only getting steeper.

it's very common for huge enterprise sales to subsidize loss leading public subscription offerings, convince individuals to love ur product so they want to keep using it at their job. obviously i don't know the details of anthropic or openai's economics on this, but i strongly expect the whole subscription thing will shake out to some minor repricings or available model shifts in a year or two, not some kind of cost apocalypse. it's pretty clear the only thing that actually affects their bottom line is R&D / model training.