Token management is part of the job now.
As a solo operator, I've had to learn it the same way I learned to manage a budget. Which tasks get offloaded to a smaller model. When to use high-thinking mode versus quick-think. Queueing up automated jobs to run during the overnight window. Balancing workloads against the 5-hour reset so I don't burn through my window before noon.
I hit limits a lot. Halfway through something and the model tells me to come back in a few hours. Drop to a smaller model to stretch the budget. Batch questions in advance so I don't waste a turn. The fact that I'm rationing at all means I'm using the thing enough for the cost to matter.
Right now we're in a subsidized window. The pricing on these tools, for what they actually do, isn't a real number. It's a land-grab number. Same playbook as Uber in 2014 — cheap rides, every trip subsidized by venture money, until the market was theirs and the prices weren't cheap anymore. The capex bill on all this AI infrastructure is going to come due, and we're all going to pay more than we do today.
Then Jensen Huang said this last month: if a $500,000 engineer isn't consuming at least $250,000 worth of tokens a year, he's going to be deeply alarmed. Half your salary, in tokens. As a productivity expectation.
That reframes things. Your value as a knowledge worker starts to correlate with how much AI you can put to work. Not how hard you work. How much throughput you command.
Same math for a solo operator, different scale. The question stops being "can I afford this tool" and starts being "what's the smallest version of me that can deploy the most tokens against the right problems."
Near term I think it gets harder. Subsidies thin out, prices go up, the cheap era ends. Long term I'm betting the other way — the compute build-out is real, local inference eats the routine stuff, frontier models get reserved for the heavy lifts, and the per-token cost collapses even as total usage explodes.
Anyway, hit the wall again this morning. Back at it in a few hours.