Every turn, an agent sends the whole conversation again. Anthropic bills the repeated part at a tenth of the price or less (a twentieth on Sonnet 5.5), but only if you switch prompt caching on with one field, cache_control. OpenAI and Google cache by default. Anthropic does not.
So I took the 718 repositories created on GitHub in October that default to a Claude model (710 still online), read every file at a fixed commit, and hand-checked 90 of them. About 147 are real agents. Roughly two in three (65.9%, 95% interval 52 to 77%) never turn caching on.
Then I ran it for real: the same 20 agent requests, twice, byte for byte. On the last request, without caching I paid full price for 54,340 tokens. With it, for 2. Input cost fell 83.1%. The whole test cost $2.39.
The lesson for anyone building agents: the default is the decision. Most people never change it.
Full article on X: https://x.com/con_golias/status/2117686000893694342
Code and data: https://github.com/con-golias/workings/tree/main/2026-11-03-claude-prompt-caching
