Btw, think this chart from today's blog has so many interesting takeaways:
- Large majority of coding agent cost is from cache hits.
- Reasoning accounts for a full 10% of cost now. When models overthink it costs a lot
- Reading files costs more than other tools combined