$ man context-wiki/model-selection

Not all AI models are the same. Some are faster and cheaper. Some are smarter and more expensive. Using the wrong model for the task wastes money or produces bad output. A capable model on a simple reformatting task is like hiring a senior architect to paint a wall. A fast model on a complex architecture task is like hiring a junior intern to design the building. Match the model to the task. That is the entire strategy.

PATTERN

Fast models (Sonnet tier) work for: reformatting content, scanning files for patterns, simple code edits, copy-paste-and-adapt tasks, and straightforward data transformations. These tasks have clear inputs, clear outputs, and low ambiguity. Capable models (Opus tier) work for: architecture decisions, complex debugging, deep research synthesis, creative content with nuanced voice requirements, and multi-step reasoning where each step depends on the previous one. These tasks have ambiguity, tradeoffs, and require judgment. The daily tracker logs my model usage so I can see where I am overspending. If I see Opus sessions on tasks that Sonnet could handle, I adjust. If I see Sonnet sessions that produced garbage, I know I needed a more capable model.

PRO TIP

When I launch parallel agents, I assign models per task. The orchestrating agent (the one coordinating everything) uses the default capable model because it needs to reason about dependencies and context. Sub-agents doing straightforward work (build a page that mirrors an existing one, update a config file, run a build check) use fast models. Sub-agents doing the heavy creative lift (writing 17 wiki entries in my voice, architecting a new feature) use the capable model. This is not about being cheap. It is about being efficient. A fast model that completes in 30 seconds on a simple task is better than a capable model that takes 2 minutes on the same task. Speed matters when you have 5 agents running in parallel.

FORMULA

Capable model sessions cost roughly 3-5x more than fast model sessions. Over a full day of building, that difference compounds. The daily tracker calculates my model usage and flags outliers. The formula is simple: for each task, ask two questions. (1) Does this task require reasoning or judgment? If yes, use the capable model. (2) Is this task mechanical or pattern-based? If yes, use the fast model. If you are unsure, start with the fast model. If the output is bad, escalate to capable. Better to try cheap and upgrade than to default to expensive on everything.