If you are going to build with AI, start now: you are being subsidised without knowing it

Recently, I wanted to know what all this really costs. Not what I pay, but what it would be worth.

The figures

My two companies have professional accounts with ChatGPT and Anthropic: four subscriptions in total, and all my bots work through them. That forces me to measure every token.

I took my actual usage over the last 30 days and applied the official API rates from Anthropic and OpenAI, checked on 25 October 2026:

• 30 days at API rates: about $12,300 (around €11,000).
• And it is rising. The average over the last week is around $900 a day. On the busiest days, between $1,000 and $1,400.
• At the current pace, that is $25,000–30,000 a month.

With my subscriptions, that same capacity for work costs me between 20 and 35 times less than I would be charged for usage, token by token. And that is a minimum: there are surcharges for very long conversations that I have not included.

By model, at API rates, GPT‑6 Astra accounts for about $4,800, Claude Opus 5.5 for about $4,000 and GPT‑6.1 Sol for about $1,800. The rest are far behind.

Probably one of the biggest subsidies in history

This is no billing error. It is a strategy. Model companies compete to attract users and businesses, and to retain those they already have. A professional subscription at a fixed price is the way in, and anyone who takes it now ends up organising their work around that tool.

How long will it last? Nobody knows. In the short term, I am not worried: the companies have an enormous incentive to sign up more customers. But an introductory price is not an equilibrium price.

The price per token will fall…

The underlying trend is downwards. Competition between the United States and China in AI models is fierce. Data centres are being built at an unprecedented pace, and there is growing talk of connecting them to nuclear power stations to secure their energy supply. Each generation of models does the same work for less.

… but usage will rise faster

In 1865, the economist William Stanley Jevons observed something that seemed contradictory: more efficient steam engines did not reduce coal consumption in England. They sent it soaring. As coal became cheaper to use, it was used for many more things. This is the Jevons paradox, and it has happened again with oil, electricity and bandwidth.

We can already see it with artificial intelligence. When working with a bot is almost free, you do not use one. You use twenty. You do not ask it for an answer: you ask it to review, compare, review again and coordinate other bots. My usage looks nothing like it did six months ago. Not because the price has fallen, but because I have found many more things to do with it.

What this means for a business

Artificial intelligence lets you do much more with fewer people, and the savings are real. But watch the numbers: those savings are calculated today at a subsidised price. If tomorrow the price per token halves but usage increases tenfold, or if subscriptions are adjusted to the actual cost of the service, replacing labour with artificial intelligence may not be as cheap as it looks now.

My conclusion has three parts:

1. Build now. Learning, experimenting and building your own system will never be this cheap again. What you learn today stays with you when prices rise.
2. Measure from day one. Record what you use, with which model and at what API rate, even if you do not pay it. It is the only way to know how much your business depends on a subsidy that will end one day, and to be ready with a choice of models when that moment comes.
3. Optimise using what you measure. Tracking your processes’ usage in detail teaches you to assign each task the right model with the minimum context needed, to automate everything that can already be automated, and to reserve frontier models for planning, delegating, monitoring and auditing. At least, that is what works for me.

Read on: Tokens: what artificial intelligence really consumes when you give it a job · The context window: the great token guzzler · I already pay for ChatGPT: why does my automation come with another bill?