Get More Model for Less: How Opus 4.5's Effort Parameter Trims Your Token Bill
If today has everyone else comparison-shopping for deals, developers running Claude Opus 4.5 have their own version of the same instinct: the new effort parameter, which sets how hard the model works on a task independently of which model you've picked. Instead of stepping down to a smaller model to save cost, you can keep Opus 4.5 and simply turn the effort dial down for routine steps.
What the numbers actually show
- Medium effort matches Sonnet 4.5's best recorded score on Anthropic's internal evals — while using 76% fewer output tokens to get there.
- Highest effort beats Sonnet 4.5's best score by 4.3 percentage points, and still uses 48% fewer output tokens than Sonnet 4.5 needed for its own best run.
For high-volume, low-stakes agentic steps — routine file edits, straightforward lookups, repetitive transformations — medium effort is worth trying first: it's already matching Sonnet 4.5's ceiling at a quarter of the output-token cost. Save highest effort for the steps where a wrong answer is expensive to unwind, like a plan that many later steps depend on.
Because effort is a request-level parameter rather than a model choice, it's also easy to A/B: run the same prompt at two effort levels against a held-out set of real tasks and compare cost against error rate before committing a whole pipeline to one setting.