“These are list-price counterfactuals, not real bills.”
I always wonder what things in the future I will hate and younger generations will be fine with. Drones everywhere and total surveillance were my guesses. Having all copy on the planet be written by poorly executing “trillion monkeys with typewriters” machines wasn’t on my bingo card. And as always, it’s the fact that all that needs to be done is to proof the copy and edit it to make it better but that lack of effort signals to me that this is low effort work and not worth my time to engage with. If it’s not worth your time, it’s not worth mine.
thank you very much for the input. I once said to my friend that "no actually you do not get $20k worth of API usage for $200 subscription, it is mostly $600". turn out it was a pretty close guess.
even if anthropic does not apply all the fancy techniques that deepseek is using for their cache system, tiered kv cache system is something that anyone implements. the US labs even have huge CXL memory pool system. basically they charge us premium for what is almost free. no wonder they are pretty generous with the quota.
inb4 "but you will have to use api if not for the sub", no it is the other way around: because you have the subscriptions so people rarely have to pay for that input cache markup. if they have to pay that much money they would just simply switch models.
It'll be fun when we have all trained the AI models and paid for it, then all realize how we have priced us all out of making money if we don't stop this unless we can afford what it will cost (monetarily and otherwise, later).
“These are list-price counterfactuals, not real bills.”
I always wonder what things in the future I will hate and younger generations will be fine with. Drones everywhere and total surveillance were my guesses. Having all copy on the planet be written by poorly executing “trillion monkeys with typewriters” machines wasn’t on my bingo card. And as always, it’s the fact that all that needs to be done is to proof the copy and edit it to make it better but that lack of effort signals to me that this is low effort work and not worth my time to engage with. If it’s not worth your time, it’s not worth mine.
I can't stand this so much that anyone who leaves this in their "writing" gets an immediate "this person has bad taste" label in my mind.
I built an e-ink dashboard for tracking my Claude Code and Codex usage, https://github.com/stared/trmnl-x-claude-codex.
It makes me aware that while I am on $200/month Claude Code subscription, I spend ~$2000 worth in tokens per week.
Let's cherish this discount while it lasts!
I suggest you make two other calculations:
1. Only calculate the output cost, skip input cost 2. Only calculate the cached input cost
You will be surprised
I don't want to make my assumptions how it is charged, or run a rather costly experiment of switching to API usage.
I use ccusage, and it already covers a very different pricing for cached input. I asked Opus 5 to cross-check that and claims it is fine.
Since you want data (it is for the last 30 days):
thank you very much for the input. I once said to my friend that "no actually you do not get $20k worth of API usage for $200 subscription, it is mostly $600". turn out it was a pretty close guess.
even if anthropic does not apply all the fancy techniques that deepseek is using for their cache system, tiered kv cache system is something that anyone implements. the US labs even have huge CXL memory pool system. basically they charge us premium for what is almost free. no wonder they are pretty generous with the quota.
inb4 "but you will have to use api if not for the sub", no it is the other way around: because you have the subscriptions so people rarely have to pay for that input cache markup. if they have to pay that much money they would just simply switch models.
It'll be fun when we have all trained the AI models and paid for it, then all realize how we have priced us all out of making money if we don't stop this unless we can afford what it will cost (monetarily and otherwise, later).
But also remember that while the API cost might be >50x the subscription cost, the actual cost to run it is much much cheaper.
It will only become cheaper as models become more efficient.
Did anyone calculate this for OpenAI?
Yes, in the linked sources, SemiAnalysis did test for Codex too: https://x.com/SemiAnalysis_/status/2064815044085318040
Similar effect, but even 70 times higher compared to Claude, 40 times.
[flagged]