The Reason AI Hasn't Replaced the Workforce Is Cost
My Claude Code logs price one month of my agents' work at $15,144. I pay a subscription. That gap is the whole story of why AI is not taking your job yet, and what changes it.
My Claude Code logs price one month of my agents' work at $15,144. I pay a subscription. That gap is the whole story of why AI is not taking your job yet, and what changes it.
Claude Code keeps about a month of session logs, so this is a snapshot, not the year. Across my two Macs, the last 30 days come to eight Claude models, 74 million output tokens and 39.5 billion cache reads. Priced at each model's API rate, that is $15,144 for the month, a run rate of about $184,000 a year. To be precise about that first number: the logs record exact consumption, and I priced it at list rates. My actual Claude access is a subscription. I did not pay $15k. That figure is the shadow: what this way of working costs if you pay per token, with no subsidy.
AI Tech Bros are proclaiming that AI is about to replace swaths of the workforce. Payscale puts the average UK senior software engineer at £61,208 a year. The shadow bill for the agents behind one developer, me, runs at roughly £140,000 a year, more than twice that. The per-model rows make it plain: the single biggest line is Opus 5.5 at $9,668 for the month, about $116,000 a year on its own. That is a salary for a model, except it does not stop at the end of the month, it scales with how hard you work it.
So who pays? Right now, partly the providers. SemiAnalysis estimates that consumer subscriptions bring in about a tenth of Anthropic's revenue but consume more than 40% of its inference compute, while the API runs at margins above 80%. Subscriptions that let a heavy user burn $15,000 of tokens a month for a flat fee are customer acquisition, carried by the API. A company that replaces a person with agents does not get my subscription. It pays the API price. That is why AI keeps not replacing the workforce at the speed the feeds predict: the intelligence is almost there, the unit economics are not. A capability that costs more than a senior salary per seat does not replace the seat. It augments whoever can afford it.
The cache line is the tell. 39.5 billion cache reads in a month: 98.2% of every token my agents touched. Most of what these models compute, they computed before. Agentic work is repetitive: the same files, the same context, the same loops, all day. Caching took about $165,000 off this month's list price; without it, thirty days would have cost nearly $180,000. We are paying datacentres, in money and in power, to re-derive yesterday's thinking, and the industry's answer so far is to discount the repetition rather than stop doing it.
The way out is the way every compute technology has gone. Everything starts big and gets smaller. Mainframes became PCs became phones. The intelligence is being born in datacentres because that is where the training compute lives, and I do not think that is where it stays. The datacentres are the cocoon, not the final product. You birth the intelligence there, then it moves to the device.
That is the product vision at Quantum Encoding Ltd, The local model does not have to be the smartest in the building. It has to be cheap enough to run every day for free, and good enough to know when it is out of its depth. One row of my app's ledger already shows the end state: a capable model running locally on Ollama, 482 million input tokens, cost $0.00.
Honest hedges. List price is not cost: the API's margin pays for training the next model, and the raw compute behind my month is a fraction of the shadow bill. But list price is what a business replacing a person would pay. And local is not frontier yet. The small driver makes mistakes the big rows catch, which is exactly why the team structure from parts 1 and 2 survives the move on-device. The loop does not change when the seats get cheaper. Build, cold review, confirm, fix. Only the invoice does.
So no, AI has not replaced the workforce, and the reason is boring and concrete: at list per-token prices it is too expensive to let it work all day. The fix is not a bigger datacentre. It is smaller models, on your own hardware, running the same loops. The cocoon is where it is born. The device is where it lives.
Token counts are from my Claude Code session logs on two machines for the 30 days to 16 October 2026, one count per API request, priced at each model's list rate as published on OpenRouter that day (Claude Mythos 5.1 at Fable 5.1's rate, which Anthropic lists at the same price; cache writes at the 5-minute rate). Pounds at about $1.32 to the pound. Salary: Payscale UK, Senior Software Engineer, 2026. Subscription and margin estimates: SemiAnalysis, reported by Crypto Briefing, 10 October 2026.