Tiered Model Strategy for Claude Code

SoloDevSage · reddit · 2026-07-19

The author codified their Claude Code model strategy into CLAUDE.md, routing tasks to different models based on complexity:

Fable: Planning and pre-merge review; planning is done in a separate session, written to docs/plans/<task>.md, followed by a /clear.
Opus: Medium-to-high complexity tasks, architecture, async/state logic, performance optimization, and unfamiliar APIs.
Sonnet: Routine implementation, like CRUD, standard components, and wiring according to existing patterns.
Haiku: Mechanical changes, like renaming, formatting, boilerplate code, test scaffolding, and doc updates.

Their core takeaway: what truly burns quota isn't the thinking itself, but the dozens of rounds of file reading, testing, and minor tweaks after planning. Thus, "planning and implementing in separate sessions" might be cheaper than keeping a powerful model on the line continuously. They also mention that using a stronger model for review catches bugs missed by the implementation model, though it's still untested if providing just the diff and spec without full context is sufficient. Finally, they ask how others choose models (per task vs. fixed strategy), how to distinguish routine from non-trivial tasks, whether sub-agents exploring codebases are worth it, and where quota is primarily consumed.

Original post →

Tiered Model Strategy for Claude Code

More from coding & agent