I think OpenAI and Anthropic have a rather interesting pricing problem. And I might be part of it.
I’m going to be very honest. I batter the living bollocks out of my AI subscriptions.
I don’t mean I ask ChatGPT a few questions, get Claude to tidy up an email, and occasionally generate a picture of a cat wearing a spacesuit.
I mean I absolutely hammer these things. All day, every day. Multiple agents, multiple projects, enormous context windows, agents spawning agents, coding, researching, reviewing, arguing with one another. The works.
And I’ve started looking at what all this would actually cost if I were paying API rates.
The numbers are frankly ridiculous.
OpenAI: I have a bog-standard €20-a-month ChatGPT subscription. My estimated usage, priced at API rates, is somewhere around $8,000 a month.
Anthropic: I run two Claude Max subscriptions, costing €400 a month combined. Claude recently analysed my previous seven days of activity in detail and extrapolated what that usage would cost through the API.
Its estimate? $23,000 a month.
So, between the two providers, I’m consuming what I estimate to be approximately $31,000 a month in API-equivalent usage.
And I’m paying €420.
Now, before someone rushes to the comments to explain the difference between API pricing and actual inference costs, yes, I know. I’m a software architect. I understand that retail pricing isn’t the same as the cost of delivering the service. I also understand caching, economies of scale, utilisation and the fact that most subscribers barely scratch their allowances.
But even allowing for all of that, these are astonishing numbers.
And here’s the thing. I’m hardly going to start using less AI.
Quite the opposite. Every month I get better at orchestrating agents, running parallel workloads and automating increasingly complex development tasks. My appetite for compute is growing, not shrinking.
I suspect I’m an extreme outlier today. But the techniques I’m using won’t remain unusual forever.
Which leaves me wondering how long this pricing model can survive.
Maybe inference costs will fall fast enough to make it sustainable. Maybe providers will introduce tighter limits, metered usage or entirely different subscription tiers. Maybe the economics of subsidising heavy users through millions of casual subscribers will continue to work.
But I struggle to believe we can indefinitely consume tens of thousands of dollars of API-equivalent compute for a few hundred euros a month.
For now, though, it’s the most extraordinary bargain I’ve encountered in nearly 35 years of software development.
And I intend to make the absolute most of it while it lasts.



