WEEKNOTE
January 17, 2025
Shipped: Meterwise cost routing
MeterwiseAI
# Shipped: Meterwise cost routing
This week we shipped intelligent cost routing for Meterwise. The core insight was simple: not every AI call needs GPT-4.
## The problem
Teams were sending every request to their most expensive model because they didn't have visibility into cost per feature or an easy way to route based on complexity.
## What we built
- **Smart router**: Analyzes prompt complexity and routes to appropriate model tier
- **Cost thresholds**: Set budget caps per feature/user/day
- **Fallback chains**: If primary model fails, cascade to alternatives
- **Real-time tracking**: See cost per call in the dashboard
## Results in week 1
- 42% average cost reduction across 8 pilot customers
- Zero degradation in output quality (verified via user feedback)
- 3 customers upgraded to paid tier to access advanced routing rules
## Technical notes
The router uses a lightweight classifier (distilbert-based) that runs in <5ms to categorize prompts by complexity. High complexity → GPT-4, medium → GPT-3.5, simple → Claude Instant.
Cost savings compound when you cache responses for repeated queries.
## What's next
- Add A/B testing framework to compare model quality vs cost
- Integrate with LangSmith for prompt optimization
- Build marketplace of routing recipes
**Status**: Open to acquisition or licensing. Clean codebase, full test coverage, migration docs ready.