[agsqwe]

thinking models produce a lot of internal output tokens making them more expensive than non-reasoning models for similar prompt and visible output lengths