The real price of AI

Part 2 of my rant about AI costs.

In part 1, I wrote about how quickly my €20 and €100 OpenAI subscriptions now get exhausted. This time I went through OpenAI's published Codex credit rates, and using GPT-5.4 as a baseline, the increase looks like this:

• GPT-5.4: 1×
• GPT-5.5: 2×
• GPT-5.6 Sol: ~1.3–1.6×
• GPT-5.6 Sol Fast: ~3.3–4×
• GPT-6 Astra: ~3.3–4×
• GPT-6 Astra Fast: ~8.3–10×

For the same raw amount of tokens, Astra Fast consumes between 8 and 10 times the credits of GPT-5.4. What I do not see in my own work is an 8 to 10 times increase in intelligence.

The switch is slowly being flipped. From my point of view, this was always the game: make AI cheap enough to become indispensable, then slowly expose the real price. €20, €100 or €200 a month made AI feel almost unlimited. Now that people and companies have built it into their daily work, the subsidies are disappearing and the real economics are becoming visible.

I think we are all being pushed towards API economics. Every task, every long context and every round of reasoning will have a visible financial cost. Companies will stop asking, "Can we use AI here?" and start asking, "Is the value created by AI here greater than what it costs us?"

One might say speed matters. I say ownership matters. When we pay a human, we do not only pay for output. We pay for judgment, context, responsibility and someone who can stand behind the work. With AI, the ownership still stays with us. We still have to review the result, understand it, maintain it and deal with the consequences.

Then there are the costs that do not appear on the invoice: technical debt and the even less visible cognitive debt. We are producing more while understanding less, and making that tradeoff without thinking enough about what it will cost later.

My prediction is that the next divide will not be between companies that use AI and companies that do not. It will be between companies that know exactly where AI creates enough value to justify its full cost, and companies that burn money because using the newest model feels like progress.

I use AI and I will continue using it. But the era where more AI automatically meant more productivity is ending. The numbers are starting to force a more honest conversation.

Are you seeing an improvement in the work that matches the increase in cost?

---

From the comments

I don't see any increase in intelligence to be honest. I strongly think we should be watching the cost and even more focus on what model does what kind of work. What I have seen is that latest models overengineer more and more. It is becoming a task for devs to cut down complexity instead of growing towards a complex system. TBH it is a bit unnatural