The rate on the pricing page is not the rate on your invoice. Output costs more than input, cached input costs less than both, and the same model is sold at four different prices through four different channels. This is a selection desk for people who pay the bill.
No signup, no email. Answer what you are building, what is non-negotiable, and what you are spending. The answer changes with the numbers — that is the point.
Every model with its input, cached-input and output price side by side, plus context, licence, modalities and whether you can self-host it. Filter by vendor, licence or strength.
Browse models →Put your own token mix in and see the monthly number, which half of it is output, and where the same traffic would land on six other models. Then test whether renting GPUs ever catches up.
Run the numbers →The same model is sold five ways, and they differ on price, paperwork and risk. Plus the eight questions procurement will ask before they let you sign anything.
Compare channels →Output tokens routinely cost 3-5x input. A model that wins on input price can lose by a mile on a workload that generates long answers. Compare output first.
DeepSeek V4.1 Flash charges $0.30/M for input and $0.006/M for a cache hit — a 50x gap. If your prompt has a stable prefix, the headline price is irrelevant.
GLM-5.3 is $1.40/$4.40 hosted by Mistral and by Together — and the two differ on cached input. Neither number is "the" price; both are real, for a specific channel.
Prices change quietly and often. Each move below is dated and linked.
How these numbers are gathered, and what we refuse to publish, is on the methodology page.