Document the core insight that working entirely on Opus is impractical
due to rate limits, and that strategic model tiering (Opus for planning/review,
Sonnet for implementation/testing) produces equivalent results with
dramatically lower rate limit consumption.
Key points:
- 90% of tokens go to mechanical tasks suitable for Sonnet
- Opus reserved for strategic decisions and pattern detection
- Batching amplifies efficiency through context amortization
- 3-5x more stories completed per rate limit window
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
VS