Start with the honest version
The case study doing the rounds is a developer who took a Claude Code spend of roughly €1,185 a month down to about €180. It is a real write-up and the mechanisms in it hold up. But read it carefully before you quote the number.
Two caveats matter. The €180 endpoint is roughly what a heavy subscription plan costs anyway, so part of the saving is a move from metered API usage onto a plan, not pure efficiency. And part of the original overspend traced back to Anthropic-side issues that were later fixed, so a chunk of that gap closed without anyone doing anything.
So: do not expect an 85% cut. Do expect the four levers below to be real, because each one is separately documented.
Lever one: stop breaking your cache
Prompt caching is the single biggest lever, and the way most people lose it is by fiddling. Anthropic’s prompt caching documentation sets out the pricing: writing to the cache costs more than a normal input token, reading from it costs a small fraction of one. That maths only pays off if you keep hitting the same cache.
Two habits destroy it. Switching models mid-session, which invalidates the cache. And adding or removing tools and MCP servers mid-session, which changes the prefix everything else is cached behind.
## Session hygiene
Do not switch models mid-session. Start the session on the model the work needs.
Do not add or remove MCP servers and tools mid-session. Decide the tool set at the start.
If I ask for something that needs a different model, tell me to start a new session instead.Watch your cache hit rate rather than trusting a feeling. Around 90% is a healthy session. If you are well under that, something in your setup is changing between turns.
Lever two: stop paying for maximum effort by default
Running everything at the highest effort setting burns roughly twice the tokens of a medium setting. That figure comes from the write-up above and has been independently replicated elsewhere at around 1.8x.
Most of your day does not need it. Reading a file, renaming a thing, drafting a first pass — medium is fine. Save maximum effort for the work where being wrong is expensive.
Lever three: route the model to the job
The cheapest token is the one a smaller model handled correctly. Put the routing rule in writing so it applies to every agent you have, not just the ones you remember:
## Model routing
haiku - mechanical work with a clear definition of done: renames, log scans, formatting, list checks.
sonnet - reading, searching, research, drafting, first-pass code changes.
opus - work where the plan itself decides the outcome: architecture, tricky debugging, reviews.
Every subagent file must set a model explicitly. Never leave it to inherit.The last line is the one people miss, and it has its own post: The Silent Opus Tax. A subagent with no model field inherits the one that called it.
Lever four: keep long work out of the main window
A long session gets expensive because the whole conversation is re-sent on every turn. Push bulk reading — scanning a directory, trawling logs, summarising documents — into a cheap subagent that returns a short answer. The main session pays for the answer, not for everything the agent read to produce it.
This is the one lever that genuinely does not change how you work. It changes where the reading happens.
What to measure
- Cache hit rate per session. Target around 90%.
- Number of sessions where you switched model mid-flight. Target zero.
- Agent files with no explicit model. Target zero.
- Share of your work running at maximum effort. It should be a minority.
Fix those four and the bill comes down on its own. No new tooling, no discipline you have to remember daily, and no dramatic before-and-after number you would have to defend.
Sources
From idea to operation
Make the next AI decision concrete.
NavAIgate helps leadership teams identify high-value AI opportunities, prove them safely and turn the winners into working systems.