So günstig wie heute wird KI nie mehr

Fable 5 came back Wednesday evening. I used it for a customer market analysis, then checked the price: $50 per million output tokens, compared to $15 on the standard model.

We will never have as cheap AI as we have today.

The task went well, and I was able to merge one of my customers' market segments with my own research into a single hybrid document.

But the pricing is what kept me thinking.

Fable 5 is the most expensive model Anthropic has ever priced for general use, double the price of Opus and triple the price of Sonnet.

The launch followed a pattern I expect to see everywhere: dumping prices first, then a nice margin.

At launch, the model was included in the paid plans. By the time I could use it this week, every request is metered at the full rate.

The big players are still losing money on every token, and the only lever they have to fix it is to raise prices.

At the same time, each task now burns far more tokens.

Two years ago, you asked a model a question, and it answered for about 2,000 tokens, costing 2 cents.

Now the same question runs through an agentic system that delegates, retrieves, and validates. Suddenly, it burns half a million tokens and costs $1.25.

Roughly 60 times the cost.

That's why I only reach for the expensive model when the task really needs it.

For the simple work, open-source models already do a great job. You don't go on Fable 5 to translate an email or summarize a PDF.

And the next price increase already has a date.

One of the current Claude models runs on introductory pricing until August 31. From September 1, the standard rate is 50 percent higher.

The visual below is from an article I came across, not my own analysis.

It puts the whole equation in one line: even when the price per token drops, you still end up paying more.