People are adopting the current models, mostly coders right now. Most people using LLMs are using them for basic information search, ie. Google AI overviews. Those tasks don't require the top models and don't burn that many tokens so the companies keep them free to get the public exposed to AI.
Then there's the 3% of people who are the power users using it for work, especially coders. For productivity the top models do perform better and the stakes are higher so they need to perform better. The top models aren't needed for every task, but they shine as an orchestrator of smaller models handling more basic tasks. Coding also requires a lot more tokens, to read all the existing code; to do "thinking" which generates a bunch of output tokens to imitate reasoning; then to generate the code itself; then to review it, adjust after a review...
All of that equates to large bills for token spend to anthropic or open AI. My Claude bill on my company account is approaching $2,000 this month, and that's about average / what's expected from an engineer at my company. I'd bet that every engineer in silicon valley is burning through a similar amount as well.
This is why both open AI and anthropic are both seeing extremely high revenue growth. Anthropic ended 2025 with $9b in revenue , they are now on track to hit $100b revenue this year, for reference / the analogy coca cola made $47b this year, it's a larger revenue then every other software company except Microsoft. THAT IS INSANE, no company has ever seen that kind of growth. Yes it's being weighed down by training costs so they aren't profitable but if they even double there revenue next year that could change.