Frontier models got cheap, open weights got good
Three AI labs shipped new models within a day of each other, prices dropped, and an idea I had shelved as too expensive suddenly cost less than lunch.
In early July, three of the biggest AI labs shipped new models within a single day of each other: Grok 4.5, the GPT-5.6 family, and Gemini 3.5 Pro after a small delay. The models were good. The interesting part was the price tag: with three rivals launching at once, nobody could afford to be the expensive one, so rates dropped across the board.
Why does that matter? Because for most of us the limit was never what these models can do. It was what they cost.
Small story. Earlier this year I had an idea: take six years of old repair reports, the kind nobody ever reads again, and let a model sort them into a simple table. What broke, how often, which part. I did the math on what processing all those pages would cost and quietly dropped the idea. Too expensive for a nice-to-have.
This month I did the math again with the new prices. The whole thing now costs less than a lunch. So I ran it. Six years of knowledge that lived only in the heads of two people, both retired by now, is now a table anyone can search. And here is the part I keep thinking about: I did not change anything. Not the idea, not the code. The world changed the price, and a “no” quietly became a “why not”.
That is the real meaning of a price drop, and it is easy to miss between the launch headlines. New capabilities change what is possible. New prices change what is worth doing. Most ideas do not die because they are impossible. They die because they are not worth it at today’s price. So I keep a little list of ideas I dropped for being too expensive, and I recheck it every time prices move. Twice this year a dead idea came back to life while I was not looking.
The second July story: open models got seriously good. Open models are the ones you can download and run on your own computer, instead of sending your data to a company’s servers. The newest batch, with names like GLM-5.2, DeepSeek V4 and Qwen 3.6, is now so close to the paid ones that in normal daily use you barely feel the difference. I spent a weekend running one on a computer that was standing around anyway and asked it my standard forty test questions. It got 36 right. For anything private or sensitive, “why are we sending this to someone else’s server at all” has become a fair question.
I also got a prediction wrong, for the record. In spring I assumed the big labs would keep their premium prices, because why would they cut them? What I underestimated was how much they fear the free alternatives and each other. You can only charge premium prices if people feel the difference in daily use. Increasingly, they do not. I have no idea how low prices will go, and I have stopped pretending to know.
One thing did not change in July: everything I built around the models kept working. The tools I connected, the test questions I wrote, the way my little assistant searches documents. Cheap model, expensive model, open model: all of it plugs in the same way. The models get replaced every few weeks. The things around them stay. If this year has one lesson for someone building things alone, it is that: put your effort into the parts that do not get replaced every month.
Oh, and the new models ship in variants named Sol, Terra and Luna. I have an open question about what those names actually mean. I suspect the answer is “marketing”.