For most of the last three years the story of the frontier models has been a story about capability. Each release was pitched as smarter than the last, and the number everyone quoted was a benchmark score. This summer that changed, and the change is worth naming, because it says something about where the whole industry is heading.
In late July the two leading American labs each shipped a new flagship. Neither led with intelligence. Both led with price.
What each of them actually said
Anthropic released Claude Opus 5 on 24 July. The claim in its own announcement is not that it is the smartest model ever made. It is that it “comes close to the frontier intelligence of Claude Fable 5 at half the price.” On one coding benchmark the company reports the new model landing within half a percent of Fable 5’s peak score while costing half as much per task. The pitch is efficiency. You can read Anthropic’s framing here.
OpenAI was even plainer. It titled the launch page for its GPT-5.6 family “Advancing the price-performance frontier”. Not the intelligence frontier. The price-performance frontier. The family comes in three tiers with names borrowed from the sky: Sol at the top, Terra in the middle, Luna at the cheap end. The whole pitch is more intelligence from every token and stronger performance per dollar.
Then they cut the prices. On 30 July, days after launch, OpenAI dropped the cost of Luna by eighty percent and Terra by twenty. Luna went from one dollar per million input tokens to twenty cents. The flagship, Sol, got a promotional rate that runs into November. A model does not get eighty percent cheaper because the compute got eighty percent cheaper in a week. It gets cheaper because the labs are fighting over who is cheapest at a given level of intelligence.
Why the pitch moved
Two things are happening at once.
The first is that raw capability is getting harder to sell. The jump from a weak model to a strong one was obvious to anyone. The jump from a very strong model to a slightly stronger one is visible mostly on benchmarks, and benchmarks are a poor way to move a customer. When two models are both good enough for the job in front of a buyer, the thing that decides between them is not the last half a percent of reasoning. It is the invoice.
The second is that the customer changed. The heaviest users of these models are no longer people typing into a chat box. They are other pieces of software making thousands of calls in a loop. For that kind of user the price per token is not a detail, it is the entire business case. Halve it and workloads that were too expensive to run suddenly pencil out. That is why both labs are selling the same intelligence at a lower number rather than a higher intelligence at the same one.
What it means
This is what the early stage of commoditisation looks like. It does not mean the models stopped improving. It means the improvement that sells has shifted from the top of the capability curve to the cost of staying on it. The frontier is still moving. It is just that the part of it the labs choose to advertise is the price.
For anyone building on top of these models, that is good news in the short run and a warning in the long run. Good news, because the same work costs less every few months. A warning, because when the underlying model is a cheap and interchangeable commodity, the value has to come from somewhere else. It comes from what you wrap around the model: the data you feed it, the tools you connect it to, the specific problem you solve better than a general chat box can.
The summer’s releases were a signal, and it was not subtle. Two of the most advanced labs on earth looked at their newest models and decided the most compelling thing to say about them was how little they cost to run. Remember that the next time a launch is described as a leap. Sometimes the real news is in the price list.
Photograph: Steve Jurvetson, CC BY 4.0.