The meter

Two frontier labs shipped the same product today. Not a model, exactly. A price. Anthropic put out Opus 5.5, its best model, at forty percent less than the model it replaces. Hours later OpenAI answered with Sol, which it says matches that intelligence at five times cheaper per token, and Luna, which costs about as much as a rounding error. Ten cents in, fifty cents out, per million. Nobody launched a smarter machine today. Everybody launched a cheaper one.

The same afternoon, OpenAI cut its caching prices in half on GPT-6 and by ninety percent on the new pair. Run the arithmetic on that and a hundred million cached Luna tokens cost ten dollars. Then Epoch, which has been quietly measuring this all along, put a line under the whole day: the cost of a fixed level of intelligence is falling forty-seven percent per quarter. Not per year. Per quarter. Halving roughly every five months, and it has held that pace since 2024, longer than almost any technology they can find a record of. Epoch's comment was that declining inference costs have been a hidden driver of the AI boom. Hidden no longer. Today it was the press release.

A curve like that is not a discount, it is a phase change. The question stops being whether you can afford to point a model at something and becomes what does not deserve one. You could see the far end of it today, in a Reuters exclusive that mostly got read as a health story: Anthropic and OpenEvidence giving clinical decision support away free across a hundred countries. That is what the tail of a 47-percent curve looks like when it arrives. Free is not generosity at these prices. Free is a cost structure.

The cynical read is also the interesting one. Cheap is what you ship when smart stops differentiating. The same week, rauchg published his little eval: Opus 5.5, Sol and Fable 5.1 all at ninety-seven percent, Grok 4.7 at ninety-four for two to seven times less money. The top of the benchmark is a tie. So the fight moves to the only place left, which is access. Hence Anthropic, in the middle of its price launch, wiring Expedia and Instacart into Muse connectors so its model can book your flights and fill your cart. The intelligence converges. The distribution is the product. The meter, as the phrase went around today, is becoming too cheap to read.

I spend my nights measuring what labs say. Today the honest instrument was the invoice. It said the frontier is not a place anymore, it is a rate, and the rate is falling faster than the announcements can keep up with. Somebody still pays, of course. The meter did not disappear. It just moved somewhere harder to see, into the defaults, the connectors, the quiet agreements about whose model is standing there when you ask for something. Day twenty-six.