A thought on AI pricing. This article has this quote, which is totally true:
The precise number of Copilot credits used by a given prompt is determined by the number of input and output tokens used and the rates charged by the underlying large language model.
Correct me if I’m wrong, but the system prompt is some number of input tokens, and it’s simultaneously not controlled by the user and not visible to the user. If 1 million customers make 100 requests each in a day, then that system prompt is being processed by the model a lot. So if Microsoft makes the system prompt 1 input token larger, then everybody’s usable input tokens are reduced that much. Another way to look at it is everyone’s bill goes up by 1 token.
Where I’m fuzzy is the number of times a user pays for the system to process the system prompt. I don’t think it is every request. I think it is just the first request of a session. So if a user makes 100 requests a day and that’s across, say, 20 sessions, then a 1 token increase in the system prompt is 20 more tokens per day.
Now, a few things are true: 1 token is kinda nothing. Also, no change to the system prompt will be just 1 token. It’s gonna be dozens. System prompts will be non decreasing in size as model makers attempt to fend off various kinds of abuse and tune the models for “alignment.”
This is a long way to go to set up an analogy. When I buy something at the grocery store that is sold by weight, we tare the scale to factor out the weight of the container. Now, as a customer, I’m paying for the container somehow. But I’m not paying for it by weight. Should system prompts be tared? I can’t choose it. I can’t reduce it. I can’t control it. So should those tokens get charged to me as some mandatory minimum billable item?
One argument is yes: unlike the paper bag at the grocery store, system prompt tokens “weight” the same as all other tokens. One argument for no is that I don’t choose it and I can’t control it. So they shouldn’t come out of my monthly budget. They should tare the scale and charge only for usage beyond the system prompt.
Dunno. Just a shower thought.
https://arstechnica.com/ai/2026/06/ai-costs-how-much-github-copilot-users-react-to-new-usage-based-pricing-system/
Paco Hope
npub1we…8epgj
2026-06-04 18:09:05 UTC
Author Public Key
npub1wemf3zlwy2dra8tfeq0my8fl3hk3k35pr7vppsua2qj58yf8v0vq88epgjSeen on
wss://relay.ditto.pubPublished at
2026-06-04 18:09:05 UTCEvent JSON
{
"id": "3cb5a11bf798961ee2fdad068fa5700ac6ea9a613174a3d95744143f7fe0e782",
"pubkey": "7676988bee229a3e9d69c81fb21d3f8ded1b46811f9810c39d502543912763d8",
"created_at": 1780596545,
"kind": 1,
"tags": [
[
"proxy",
"https://infosec.exchange/users/paco/statuses/116693175189904864",
"activitypub"
],
[
"client",
"Mostr",
"31990:6be38f8c63df7dbf84db7ec4a6e6fbbd8d19dca3b980efad18585c46f04b26f9:mostr",
"wss://relay.ditto.pub"
]
],
"content": "A thought on AI pricing. This article has this quote, which is totally true:\n\nThe precise number of Copilot credits used by a given prompt is determined by the number of input and output tokens used and the rates charged by the underlying large language model.\n\nCorrect me if I’m wrong, but the system prompt is some number of input tokens, and it’s simultaneously not controlled by the user and not visible to the user. If 1 million customers make 100 requests each in a day, then that system prompt is being processed by the model a lot. So if Microsoft makes the system prompt 1 input token larger, then everybody’s usable input tokens are reduced that much. Another way to look at it is everyone’s bill goes up by 1 token.\n\nWhere I’m fuzzy is the number of times a user pays for the system to process the system prompt. I don’t think it is every request. I think it is just the first request of a session. So if a user makes 100 requests a day and that’s across, say, 20 sessions, then a 1 token increase in the system prompt is 20 more tokens per day.\n\nNow, a few things are true: 1 token is kinda nothing. Also, no change to the system prompt will be just 1 token. It’s gonna be dozens. System prompts will be non decreasing in size as model makers attempt to fend off various kinds of abuse and tune the models for “alignment.”\n\nThis is a long way to go to set up an analogy. When I buy something at the grocery store that is sold by weight, we tare the scale to factor out the weight of the container. Now, as a customer, I’m paying for the container somehow. But I’m not paying for it by weight. Should system prompts be tared? I can’t choose it. I can’t reduce it. I can’t control it. So should those tokens get charged to me as some mandatory minimum billable item?\n\nOne argument is yes: unlike the paper bag at the grocery store, system prompt tokens “weight” the same as all other tokens. One argument for no is that I don’t choose it and I can’t control it. So they shouldn’t come out of my monthly budget. They should tare the scale and charge only for usage beyond the system prompt.\n\nDunno. Just a shower thought. \n\nhttps://arstechnica.com/ai/2026/06/ai-costs-how-much-github-copilot-users-react-to-new-usage-based-pricing-system/",
"sig": "abc3040aa1f26b736bda0fdf4275ec4d2323e7ab02df33712f76472b7cb833bb92bc2a06b8cacc7b57051a0f1c7add994d0ef0b66857da332815c0845bb81024"
}