i am in a similar position with
@the_smart_ape


i've burned through a $200 Anthropic API subscription in 2 weeks without doing anything crazy

i realized i fundamentally misunderstood where tokens actually go

tokens burnt ≠ what you type + what AI responds, that's maybe only 20% of it

the other 80% is invisible, full system prompt, context files, entire conversation history resent to the model on EVERY single message

on Opus that's $0.15-0.22 per message before you say a word

here the hack that fixes costs by 80%:
→ slim context files brutally
→ Sonnet as default. Opus on demand only. Costs drop 65% immediately
→ move instructions into skills. description = 20 tokens. Full instructions load only when needed
→ compact at 50-60% context.
→ cron jobs on Sonnet.
→ watch your cache hit rate. <50% means you're overpaying massively

@AnthropicAI killing my wallet every month and this is a good hack

recommend studying this

https://x.com/arndxt_xo/status/2029404900573258020