Lesson 07: Model and Effort
Two frontmatter fields let you choose a different model and reasoning effort level for a specific skill, independent of the session defaults.
The model field
model: haikuValid values: haiku, sonnet, opus, a full model ID string (e.g., claude-opus-4-5), or inherit.
inherit is the default — the skill uses whatever model the current session is running on.
When to override the model
Use haiku for tasks that are fast, well-defined, and do not require complex reasoning:
- Formatting a file according to a style guide
- Extracting specific fields from structured data
- Generating boilerplate from a template
- Simple one-shot transformations
Use sonnet for most coding tasks. This is the default for Claude Code bundled skills because it balances capability and speed well for the majority of development workflows.
Use opus for tasks that genuinely require extended reasoning:
- Architectural analysis or design decisions with significant tradeoffs
- Security audits where missing a subtle vulnerability has real consequences
- Root cause analysis for complex bugs where previous analysis attempts have failed
- Reviewing major changes to core infrastructure
Cost tradeoff
These are illustrative multipliers, not exact pricing. Actual pricing varies by model generation and should be verified at anthropic.com/pricing:
| Model | Relative cost | Best for |
|---|---|---|
| Haiku | ~1x | Formatting, extraction, boilerplate |
| Sonnet | ~5x | Most coding tasks (session default) |
| Opus | ~25x | Architecture, security, complex debugging |
A skill that uses model: opus costs roughly 25 times more per invocation than a haiku skill. For a skill you run once for a serious security audit, that is worthwhile. For a skill you run after every file edit, haiku is almost certainly the right choice.
The override is temporary
The model override applies only to the skill’s turn. The next prompt in the session uses the session model, not the skill’s model.
The effort field
effort: highValid values: low, medium, high, xhigh, max.
Effort controls how much extended reasoning the model applies before responding. Higher effort takes longer and costs more but produces better results on tasks that benefit from deeper thinking.
When to override effort
Use effort: low for tasks where the answer is immediate and mechanical: a simple string transformation, a yes/no check, a code formatting pass.
Use effort: high or xhigh for tasks that benefit from the model considering multiple approaches before responding: architecture decisions, security analysis, optimizing a complex algorithm, or any task where the session default produces results that miss important considerations.
Use effort: max sparingly. This is the equivalent of asking the model to think as hard as it can. Reserve it for genuinely difficult problems where lower effort has already produced unsatisfactory results.
Combining model and effort
The two fields are independent:
model: opuseffort: highThis runs the most capable model with extended reasoning — appropriate for a security audit or a hard architectural question. It is also the most expensive combination.
model: haikueffort: lowThis is the fastest and cheapest combination — appropriate for a formatting pass or a simple code generation task.
Examples
examples/deep-analysis/— usesmodel: opusandeffort: highfor architectural analysis and security reviews