Two dollars and ten dollars. That is what a million input and output tokens on Claude Sonnet 5 cost this week, roughly a third below its own standard list price that starts in September. On the same days, Meituan released LongCat-2.0, a 1.6-trillion-parameter model under an MIT license that edges GPT-5.5 on a coding benchmark and costs nothing to license. The floor under applied AI is being poured from both directions at once.
Here is the mechanic. Intro pricing buys adoption before switching costs exist: land a workload on Sonnet 5 at $2/$10 now, and the $3/$15 standard later meets a customer who has already built around it. Open weights attack from below, resetting what a buyer will pay a closed lab once a free model does the job well enough.
For an operator the spread is the whole story. Our support-assistant workload runs from $113 a month on open-weights APIs to $1,900 on frontier models, a 17x range for the same 100k messages. That gap is not a price to accept. It is a design decision, and this week both ends of it moved.
Support-assistant workload holds a 17x spread: $113/mo open-weights floor to $1,900/mo frontier ceiling for 100k messages.
