First, measure
Without measurement, any cut is made in the dark. cognitive-claude includes a tool that reads your local session logs and shows cost per turn, cache usage and the work and cost split by model:
python3 tools/cost-audit.py --window 90 --verboseWhat the numbers showed
90-day window, Jan 29 to Apr 28, 2026, 987 sessions:
| Model | Work | Cost |
|---|---|---|
| Opus | 52,5% | 94,1% |
| Sonnet | 19,9% | 4,3% |
| Haiku | 27,6% | 1,6% |
Almost half the work was already done by smaller models, for less than 6% of the cost. And 71.76% of the work was done by sub-agents, each with its own context.
Route by type of task
The logic is simple: hard decisions and planning stay on the bigger model. Mechanical tasks, like searching, reading and formatting, go to the smaller ones. Before changing a rule, compare result and cost for that type of task in your own logs.
Automate the rules
A rule that depends on remembering does not last. In cognitive-claude, about 100 lines of working rules live in a constitution file, and six hooks apply those rules automatically in every session.
What this number does not prove
This is one operator's data, and yours will differ. The README itself separates what the framework contributes from what is just the difference between plan pricing and API pricing. That is why the recommendation is to measure your own usage before copying any rule.
Frequently asked questions
Do I need cognitive-claude to do this?
No. The principle applies to any setup: measure, route and automate. cognitive-claude is an open implementation you can use or study.
Does a smaller model make results worse?
It depends on the task. That is why measurement comes first: compare result and cost by type of task before changing the rule.
Want to apply this to your operation?
The first call is a free one-hour diagnosis.