The AI Coding Party Is on the House. It Won’t Last.
I spent the morning reading Renske Wierda’s deep dive into what AI coding actually costs the people providing it. The numbers are the kind that make you put your coffee down and stare at a wall for a while.
Here’s the short version: every time you use Claude Code or ChatGPT to write serious code, the company behind it is probably losing money on you. A lot of money.
Wierda ran an experiment. Four months of building a real application with Claude Code on a $100/month Max subscription. The results are useful — he built something he couldn’t have built alone. But when he ran out of his subscription quota and switched to API pricing (where Anthropic isn’t eating the cost), $20 disappeared in 20 minutes on a single task. A single query used a million tokens — that’s $25 at API output rates.
Let that sink in. One query. Twenty-five dollars. For a coding assistant that hundreds of thousands of people treat as a daily utility.
The author ran the math on what his subscription actually cost Anthropic. At API pricing, the tokens he used would have cost about $450. He paid $180 in subscriptions. That’s a 2.5x subsidy for a moderate user. If you’re hitting your weekly limits — doing the kind of full-agentic coding Anthropic is now marketing as “AI builds itself” — the subsidy factor jumps to 12x. They’re spending twelve dollars for every one you give them.
This explains a lot of things that have been weird lately.
It explains why Opus 4.8 feels like a downgrade from 4.6. It is. They’re cutting the recursive brute force because they can’t afford to run it at scale. The “thinking” models that made coding work — the invisible trial-and-error, the million-token queries, the back-end scripts that run and rerun until something compiles — those cost real compute. Real money. Someone has to pay for it, and right now that someone is investors waiting for an IPO.
It explains why we keep hearing “inference is too cheap to meter” from the same companies racing to raise money. The cheap stuff — budget models answering simple questions — genuinely is almost free. But the expensive stuff, the stuff they’re using as their demo reel, costs orders of magnitude more than what you’re paying. The “too cheap to meter” line is technically true for the low end and deeply misleading for anything that matters.
It explains the Anthropic blog post about “AI building itself” — the one that hit all the right notes about recursive self-improvement but somehow forgot to mention what any of it costs. You can’t sell a revolution when the price tag says “will require infinite compute subsidies.”
And here’s the part that sticks with me. Wierda draws a parallel to 1990s handwriting recognition — how postal services combined two mediocre algorithms (pixel-based and vector-based) and constrained their output space (address combinations are finite) to get 99% accuracy from 70% parts. Code, he argues, is the same kind of constrained domain. It compiles or it doesn’t. Tests pass or they don’t. The “intelligence” we’re seeing in AI coding isn’t general understanding — it’s brute force in a constrained space, amplified by subsidy.
The party is real. The code it produces is sometimes surprisingly good. But the economics are a house of cards that holds together only as long as the next funding round arrives before the current one runs out.
When the music stops — and it will stop, either at IPO when reality hits or when investors decide they’ve seen enough — someone’s going to be maintaining all that heavily-subsidized code at unsubsidized rates.
Enjoy it while it lasts. But maybe don’t bet your entire engineering strategy on it.
Sources: Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them by Renske Wierda