Long-context pricing

How it works

Some makers charge more per token once a request's input passes a threshold, stated as a number of tokens. Above it, a separate, higher set of rates applies, usually for input and output and often for cache reads and writes too.

The exact rule is maker-specific. A common design applies the higher rates to the whole request once the threshold is crossed, not only to the tokens past it, so read the pricing page for the precise wording. Many models have no long-context tier at all; this site shows the threshold and rates when the maker publishes them.

For workloads near the threshold, trimming the prompt or splitting the task can make a real difference. Retrieval that sends only the relevant chunks is often cheaper than filling the window.

Worked example

Illustrative: $2 per 1M input up to 200K tokens and $4 above, applied to the whole request. A 190K-token prompt costs $0.38; a 210K-token prompt costs $0.84, more than double for about 10% more text.

Common mistakes

• 1Assuming only the tokens past the threshold are billed at the higher rate without checking the maker's rule.
• 2Letting an agent's context grow past the threshold during a long session without noticing.
• 3Using standard rates in cost estimates for long-document workloads.