Docs Optimize Open models

Open models

See what your tokens would cost on open models such as Qwen or Llama.

Every figure, and how it is computed

Spend that has an open alternative: Spend on token-priced AI calls whose model has at least one open-weight model priced in the rate card.
The same tokens, priced open: Each row on its cheapest open model at the host's published price, same tokens. Assumes nothing about quality.
Backed by a benchmark: The same, counting only models a benchmark in the register says hold up for that kind of work.
The same model, more than one price: One open model sold by several hosts. Host prices often differ more than model prices.
Or run the model yourself: The volume where your GPUs cost the same as the bill. You set price, count, speed, how busy.
It leaves out: Engineering, evaluation, reliability, storage and egress. Named, never priced.