When you swap to the cheaper model, how much does the per-task bill actually drop once the CLAUDE.md tokens are counted? Curious whether the instructions eat much of the savings.

Show Post

i guess that was ampcode who did this first and later adopted by cc

Show Post

they targeted those repeat token cost. instead of dumping the whole tool catalog into context and returning raw tool output every turn, the agent writes code in a sandbox with Bash, Python, and SQL, so only trimmed results come back.

Show Post

Nice! Those are perfect jobs for a small, local model. What type of tasks do you end up running on cloud LLMs?

Show Post

This is exactly where local wins: narrow, repeatable jobs. A small model that classifies well in under a second beats a big cloud call on cost, latency, and privacy every time.

Show Post