Update: this inspired a note on my blog: https://simonwillison.net/2026/May/29/anthropic/
This one client, then, is 12.8% of Anthropic's run-rate revenue? That does not exactly fill me with confidence that it's a meaningful number. Doesn't this suggest run-rate revenue will fall off a cliff if Anthropic customers start applying cost controls?
If they really did count that one customer as $6B it means they've gone from $30B in April to a mere $41B in May.
im really not sure why he keeps parroting on about this. its all irrelevant frankly. companies play games. non-gaap revenue recognition, adjusted operating income.... boo-ya.
wait for the somewhat official doc's to come out, then its worth talking about.
As an anecdote, Github is changing their copilot plan to usage based billing next month. They released a tool that allows their users to estimate what their bills will look like under the new plan based on their past usage. There are some screenshots online from users showing their plans will go from $40/month to $3-5k/month. I imagine this is happening everywhere. These tools absolutely can do more than they were capable of just six months ago. But if the true costs are as high as it appears, folks are going to be much more judicious with how they use them moving forward.
People confuse price with cost all the time. The price of Opus has dropped from $75/1M output tokens to $25. That's the price. The cost is much lower and according to Dario, about a month ago, they had about a 73% margin.
I don't understand how anyone would use GitHub copilot...it's basically running a custom harness and using close to API pricing for Opus. This is why Microsoft is cooked in this game.
But yeah, I don't understand why people switch from subscription to API prices for Claude. They're way higher, but again that's price and not necessarily the cost to Anthropic to serve.
Do you have a source for this? I missed this and it does not match my impression.
And FYI, most enterprise accounts were forced to switch to a hybrid monthly seat license plus API based usage earlier this year. So that is why we are probably seeing so much alarm over Ai bills at the enterprise level. Companies went whole hog on agentic workflows not fully appreciating the costs structures of their new plans. Didn't help that pretty much every VC and board was probably breathing down their neck that if they didn't jump on the AI bandwagon they would get left behind.
There were a lot of people (here included) that were absolutely abusing github - getting Opus to generate its own subagents for days on end with a single premium request. If THAT'S $3k/mo, I'm honestly not worried.
I was just asking sonnet, 5.4, opus, in single agent session to fix a problem for 20 minutes...if THAT'S $3k/mo, then AI is truly cooked.
I think for your use case, the most likely outcome is that you are going to need to be on a $100-200 month plan if you want access to the cutting edge models. But on the other end of the spectrum, you could probably get closer to $10-20 month using Chinese models. My usage was closer to yours, with the occasional tokenmaxxing sprint just to experiment a bit and test out the limits of the plan. I am not quite sure what I will be doing next month once it moves to API billing but I suspect I will move to openrouter and one of the open source CLI harnesses.