model-selection-guide

Model selection depends on task complexity, quality requirements, and token budget. Opus handles complex reasoning like architecture design; Sonnet balances quality and speed for implementation; Haiku minimizes cost for simpler tasks like test generation. Match the model tier to your task's demands.

Allocate tokens strategically using a target distribution: 20% to planning, 60% to core execution, 20% to review and refinement. Assign Opus to high-complexity phases, Sonnet to standard implementation, and Haiku to repetitive or low-risk tasks to maximize efficiency.

Enable thinking mode for complex reasoning tasks requiring deep analysis, such as security reviews or architecture decisions. Extended thinking consumes more tokens upfront but prevents costly model switches and rework, delivering better results on challenging problems where quality justifies the cost.

Architecture design and security analysis need Opus; code implementation and reviews fit Sonnet; test generation and documentation use Haiku. This phase-based mapping ensures you apply sufficient model capability where it matters while controlling costs on routine work.

Code review complexity determines the model. Use Opus for security-critical or architectural reviews requiring deep reasoning; use Sonnet for standard quality checks; use Haiku only for simple style or formatting reviews. Match model capability to review scope and risk.

Underestimating complexity wastes tokens on multiple resubmissions and model escalations, increasing total cost. Overestimating burns unnecessary budget. Model selection guidance prevents both by clarifying task-to-model mapping, escalation rules, and contextual awareness checks before you commit tokens.