A one-word answer from Claude Code cost 56 times more on default settings than with tools off

We'd taken the model out of Port42, and the obvious way to put it back was to shell out to claude -p. So we measured a one-word answer first.

With tools and settings off it used 298 input tokens and cost $0.0005. On default flags it loads the whole agent harness, about 23,000 tokens, and the first call cost $0.029. An immediate second call hit the cache and cost $0.0032.

Each call is also a new process, around 290 MB at peak. And on defaults it's a full agent with every built-in tool switched on, including shell commands. That's a much bigger permission than "use AI".

We didn't build it. A port that wants a model posts in its chat and a companion answers.

claude -p "Reply with exactly: OK" --model haiku
lean     in 298     cost $0.000513
default  cc 22,951  cost $0.028899   (first call)
default  cr 22,951  cost $0.003220   (second call, cached)
max RSS 301,858,816 bytes (lean)