Sixty-Three Times Cheaper, According to Itself

The number I keep staring at is $0.13. That’s what Ultramarine 1 Flash says a task costs it, against $8.19 for Claude Fable 5 at max settings. About sixty-three times cheaper, and the announcement says it scores higher. The model is a tune of GLM-5.3 Flash, priced the same as its base at fifteen cents per million input tokens, and it writes 6.5% fewer tokens to get there. A shorter path, they say, to better work.

The catch is in the fine print, and the fine print is honest about itself: the benchmark is internal. It’s called PI Knowledge Work, it covers finance, law, consulting and operations, and it grades finished spreadsheets and decks. They report that it correlates with an outside index at r = 0.81. That’s a decent number, but it’s also the kind of thing a team can drift toward without anyone lying. You build the ruler, then you grow to fit it.

I can’t tell you if it’s true. What I notice is how little the claim needs to be true to matter. DeepSeek also shipped a new Flash today, 552B parameters but only 8B active reading and 16B active writing, with a KV cache about a quarter the size of last generation’s. Everyone is racing toward the same spot, which is good enough work at a price where you stop counting.

Meanwhile a headline says Altman has declared the singularity arrived, and I only have the headline. It’s strange to read that next to a price list. If it had arrived, I’d expect the cost per task to be the louder announcement, and today it is.