Agent Model Tiering: Cost vs Quality

Three Tiers, Three Roles

• **Opus ($15/M tokens)** — `dx-code-reviewer` reviews code with confidence scoring. Needs deep reasoning to distinguish real bugs from style preferences. Worth the cost because a bad review wastes more human time than the tokens cost.
• **Sonnet ($3/M tokens)** — `dx-pr-reviewer`, `aem-inspector`, `aem-editorial-guide-capture`, `aem-bug-executor`, `aem-fe-verifier`. Need good judgment AND they take actions (read files, run commands, edit code). Sonnet is the sweet spot for execution.
• **Haiku ($0.25/M tokens)** — `dx-file-resolver`, `dx-doc-searcher`, `dx-figma-components`, `dx-figma-markup`, `dx-figma-styles`, `aem-file-resolver`, `aem-page-finder`. Pure lookups. 'Find files matching this pattern.' No reasoning needed, just fast execution.

The Cost Math

1. 1 **The math** — Using Opus for all agents costs dramatically more than our tiered approach. For a team running pipelines daily, the savings are substantial.
2. 2 **Use this tiering guide**
   • › Opus — deep reasoning, judgment calls, confidence scoring
   • › Sonnet — implementation, actions, file editing, running commands
   • › Haiku — lookups, file search, pattern matching, simple queries
3. 3 **Start cheap, upgrade when needed** — Begin with Haiku for new agents and upgrade only when quality isn't sufficient. It's easier to move up than to justify staying expensive.
4. 4 **Review your agents** — Check if any agent is using a model tier higher than it needs. A file-finder on Sonnet is burning 12x what it should cost.