Price a workload, not a token

Your token mix

Every number is yours to change. If you don't know your mix yet, start from one of these shapes — they are illustrative starting points, not measurements.

5.70M tokens in total · on the cheapest model, deepseek-v4-flash, cache reads and writes are 41% of the bill

What it costs, cheapest first

Prices come from the live billing catalog — the same one that bills your requests. “Official” is the provider's published list price where they publish one; a blank cell means they don't. Models billed per request or per image are not in this table because a token mix cannot price them.

Why the cache lines matter

Agentic coding tools re-send the system prompt, the project context and the conversation history on every request. Providers cache that context, so it bills at the cache-read rate rather than the input rate — which is why a session can be 80–90% cache reads by token count while the headline input price barely touches the bill. A discount that applies only to input and output, and not to the cache lines, is worth far less than it looks. The discount badge analysis works one example through end to end, and what a real Claude Code session costs breaks a single session down line by line.

On KoRouter one multiplier applies to every price component of a model, cache reads and cache writes included, and the provider's official price sits next to ours in the model marketplace so you can check the arithmetic yourself.