How We Found Over $1,000 Per Month in AI Cost Savings Without Downgrading Anything

Running AI at scale gets expensive fast. At Balay ni Bruno & Co., we hit a point where our Claude API spend was climbing every month as we added more AI Family members, more subagents, and more knowledge base files. We were not wasting money on the obvious things. We were wasting it on invisible things. So we stopped building new things for one week and ran a proper audit. What we found surprised us, and it cut our projected monthly spend by $1,000 to $1,900 without removing a single capability.

The Five Moves That Changed Our Numbers

The biggest single win was prompt caching with a one-hour TTL. Claude's default cache timeout is five minutes. We did not know this had silently changed. In March 2026, the default regressed from one hour down to five minutes, which meant our large knowledge base files were being re-read and re-charged on almost every call. The fix was simple: explicitly set ttl:1h in every request that pulls from a KB file. That one change alone is projected to save $400 to $700 per month because a KB file read once gets served from cache for the rest of the hour instead of being billed again every five minutes.

The second move was adding an MCP compressor wrapper. Tool results from Model Context Protocol calls can be verbose. Raw API responses, full JSON objects, long content blocks. All of that lands in the context window and gets billed as input tokens. A lightweight compressor that strips unnecessary fields and summarizes tool output before it enters the context window reduces that overhead on every single tool call. The third move was switching to the Files API for knowledge base retrieval. Instead of loading entire KB files into each conversation, we reference them by file ID. Claude retrieves only what it needs. This is especially impactful for our longer AI Family system prompts and client knowledge bases.

The Subagent Problem Nobody Talks About

The fourth move was the model cascade. Not every task needs Sonnet. Text edits, quick lookups, simple formatting, status checks: these run fine on Claude Haiku at a fraction of the cost. We mapped every subagent and routine task in our system and routed anything that did not require reasoning to Haiku first. If Haiku flags it as complex, it escalates to Sonnet. That cascade alone is projected to save $300 to $500 per month. The fifth move was the subagent audit, and this one was humbling. Every subagent we fire inherits the full parent context by default. That means a 50,000-token context window gets passed to the subagent, which then generates its own output, and the whole chain multiplies. We found subagents running with context they did not need. We stripped each one down to only the context actually required for the task it handles.

The Opus 4 context trap also caught us. The marketing says one million tokens of context. What it does not lead with is that the system auto-compacts at 200,000 tokens, not one million. If you are building workflows expecting to hold a full million tokens without compaction, you will get unexpected behavior. We learned to plan around the 200K effective ceiling, not the advertised max. The audit took about a week of focused work across our system. The results are now documented in our internal operations knowledge base so every new AI we build starts with these defaults baked in. At Balay ni Bruno & Co., the goal is not to spend less on AI. The goal is to spend right on AI so the budget goes toward capabilities that actually compound the business.