Your $100 AI Coding Plan May Cost the Vendor $1,000. The Developer's Read.
on june 7 gerben wierda published a piece with a blunt title: anthropic and openai may be spending more than $1000 for every $100 you pay them. he is not talking about the chat box. he is talking about agentic coding, the mode where the model reads your repo, edits across files, runs the tests, and burns through tokens with almost no human in the loop. i build small dev tools on top of this stack and i pay these subscriptions, so a claim that the thing under my whole workflow is sold below cost is not a headline i scroll past. here is the read for people who ship on top of it.
Where the number comes from
wierda's case is an estimate built from public api rates, and he says so up front. he prices opus at $5 per million input tokens and $25 per million output tokens, then walks through what real coding use does to those meters. a single million-token query against a large repo runs about $25. one complex multi-file task he resolved cost in the neighborhood of $65 in tokens. over a four-month project he paid a $100 subscription plus about $180 in overflow, and estimates the tokens he consumed would have cost roughly $450 at api rates. one bad afternoon he watched $20 drain in twenty minutes once he tipped past the subscription cap.
from that he lands on a subsidy factor: about 2.5x for ordinary coding use, climbing toward 12x for someone who pins a max plan at its weekly ceiling. that is the path to the $1000-for-$100 line. i cannot audit anthropic's books and neither can he, so treat the exact multiple as an informed guess, not gospel. but the direction is hard to argue with, and two of his mechanics match what i see on my own invoices.
The two mechanics that bite quietly
the first is caching, or the lack of it. prompt caching only pays off when calls come close together. a human in the loop leaves ten to thirty minute gaps between turns while you read the diff and think, and the cache goes cold in between. so the human-paced workflow, the one most of us run, is the worst case for the meter. i wrote about the same effect from the cost-cutting side in the three places a claude code bill leaks; uncached prefixes were the biggest of the three.
the second is thinking tokens. high-effort reasoning can spend tens of thousands of tokens per turn that you never see in the output, billed at the output rate. you feel the latency and the bill, but the work is invisible, so it is easy to assume a quiet turn was a cheap one when it was the opposite. between cold caches and hidden reasoning, the gap between what a request looks like it cost and what it cost is wide, and it points one way.
Why this is a strategy story, not a billing tip
wierda's conclusion is that the brute-force code-editing party cannot last at these prices. i agree, and i want to be precise about what that does and does not mean. it does not mean the tools go away. the coding-agent category is real and it is large enough to fund a near-trillion-dollar company. what it means is that today's price per unit of work is partly a customer-acquisition number, and acquisition pricing has an expiry date. when the subsidy compresses, it reaches your invoice as some mix of higher prices, tighter weekly limits, more aggressive routing to smaller models, or metering that used to be free.
github copilot already gave a preview of the shape of this when it moved to usage-based billing on ai credits and power users started projecting bills several times their old flat rate. the flat-rate era is the subsidized era. the metered era is the one the unit economics support. assume you are living in the first and budgeting for the second.
What i would do with this
nothing dramatic, and nothing that requires switching vendors today. but a few habits buy cheap insurance against the day the price moves.
keep your workflows portable. the value should live in how you drive the tool, your claude.md rules, your subagent and review patterns, not in one vendor's current price card. a workflow that survives a model swap survives a price change too.
measure your own token spend per unit of work, not per month. the monthly number tells you nothing about which tasks are efficient and which are quietly expensive. if you cannot see that a particular kind of task costs 10x what a similar one does, you cannot fix it, and you will not notice when a pricing change makes that task uneconomic. this was the exact blind spot i kept hitting on my own bill, so i started building tokenwise to surface where the tokens go. it is early and validating before i build the whole thing, so it is a waitlist right now; if watching an ai bill rise with no clear reason sounds familiar, that page is where to follow it.
and treat the cheap period as a window to build the skill, not a permanent subsidy to lean on. the developers who come out ahead when prices normalize are the ones who used the subsidized years to get good at driving the agent well, so they need fewer, tighter turns to reach the same result. that skill keeps paying after the discount ends.
Sources
- gerben wierda, "anthropic / openai may be spending more than $1000 for every $100 you pay them" (subsidy estimate, opus rates, four-month project figures): ea.rna.nl