ccusage
Inspired by a recent Simon Willison post, I just ran ccusage on my laptop and learned that over the past thirty days, on my sole subscription of $200 a month for Claude, I have consumed $2,422.13 worth of tokens. All of this elides the fact that ccusage anchors onto something that is itself slightly lossy: we have to take as given the assumption that the unit cost of tokens through metered billing for companies like Anthropic is, in and of itself, profitable. I suspect that it is. But it's worth calling out, since the entire $2,422.13 number rests on it.
Talk about getting my money's worth.
It is interesting — and a little bracing — to think about the hypothetical world in which no prosumer subsidy exists, and I would have to actually pay $2,422.13 to receive the output of these tokens.
First off, I think my usage patterns would change dramatically. The vast majority of these tokens are spent on Buttondown, and at that price point Anthropic would be the second-largest vendor on the Buttondown books, behind only Stripe. And this doesn't even include anyone else on the team. I would imagine that my actual spend would, in that world, dwindle to a third or a fourth of its current size — not because the marginal cost outweighs the marginal value, but because there is simply so much low-hanging fruit. I am generally in the business of saving $1,000 a month if it's easy to do so.
Two other notes prompted by Simon's essay.
One: LLM spending, at least from the outside, feels like a bit of a bokeh. I know many companies have grown much more sophisticated about this in the past four years, but during my time at Stripe and Amazon, a lot of the efficiency work "Efficiency" being the buzzword used to mean, roughly, we would like to lower our OpEx in preparation for either the next quarterly earnings report or the next round of layoffs. was not really spent doing fancy backbreaking things — it was spent figuring out which fleets of servers were simply collecting dust because some random team had turned them on six months ago and never spun them down. I joke a lot that my single most meaningful contribution to Amazon was saving us tens of millions of dollars a year because I had the great fortune of realizing one of our ad-hoc clusters for one-off jobs was not scaling down, and we were therefore burning a huge amount of money for no reason in particular.
LLM spend, at the org-chart level, smells identical to me. Distributed, badly telemetered, growing fast enough that no one at the top has had time to build intuition for what right looks like.
Two: my already fervent interest in local LLMs — and getting to the point where I can run some of the more recent engineering-grade models locally — would, in that hypothetical world, become the single highest-leverage thing I could do from a financial standpoint. The calculus on local inference is not just an aesthetic or moral preference (see uber-for-tokens); it is also an economic one, and one that I suspect will become increasingly load-bearing.