The 80% rule: when Haiku beats Opus

Default to the biggest model? That's expensive.

It's the path of least resistance: pick the most capable model, never think about it again. But capability you don't need is cost you don't need.

Where smaller models match

For these task types, mid-tier and fast models often match flagship quality - check the advisor's confidence scores for your task type rather than taking it on faith:

• -Summarization - Sonnet and Gemini Flash often handle this as well as Opus
• -Code review - catching bugs and style issues doesn't need the biggest model
• -Classification - routing, labeling, and categorization are well-solved at every tier
• -Simple generation - boilerplate, templates, and standard patterns

Where flagship models matter

• -Complex multi-file reasoning - understanding interactions across a codebase
• -Architecture decisions - tradeoff analysis and system design
• -Novel problem solving - tasks with no clear pattern to follow
• -Long-context synthesis - pulling insights from 100k+ token contexts

The 80% rule in practice

The 80% is the price gap, not a usage statistic: Claude Haiku 4.5 lists at 80% less per token than Claude Opus 4.8. Look at your last 100 API calls and sort them into the two lists above. Every call in the first list that you are paying flagship prices for is a candidate to move down a tier.

Try the cheaper model on a sample first and compare the outputs. For routine work the difference is often hard to spot; where it is not, keep the flagship.