Prompt caching for Claude sessions
Claude sessions now use prompt caching, so the conversation history in each turn is read from cache at a fraction of the input price, and usage counts that.
A session's agents work in a loop: call the model, run a tool, send the result back. Each turn used to resend the whole conversation at the full input price.
Requests to Claude now carry Anthropic's cache markers, on the system prompt and on the conversation so far. On the next turn, the history comes back as a cache read, which Anthropic bills at a small fraction of the normal input price.
We also count usage the way it is billed. Fresh input tokens count in full, and cache writes and cache reads count at their own rates. Without this, the saving would not show anywhere: not in the cost graph, not in the spend caps, and not in your monthly tier allowance. A long Claude session now uses less of your allowance than it did before.
This applies to Claude on the platform's models and to workspaces that connect their own Anthropic key. See models for the tiers.