Skip to main content
On the benchmarks we run, Agent API effort levels deliver more quality per dollar than Sonar: the effort level curve sits above every Sonar model. The lead is widest on hard, multi-step browsing: on BrowseComp, the strongest effort level scores nearly three times the best Sonar model.

​
What we measured

We evaluated Agent API effort levels and Sonar models on the same workloads, across two benchmarks:
  • BrowseComp measures agentic browsing on hard questions that require chaining many searches.
  • DSQA (DeepSearchQA) measures answer quality on deep-search questions.
Even the low effort level beats Sonar’s best model, Deep Research, on both benchmarks at 1-2% of its per-request cost. Sonar is retired and has migrated to the Agent API. The Sonar numbers on this page are the baseline those models were measured against before retirement.
For a benchmark built specifically for research agents that must search both wide and deep, see WANDR.

​
Which effort level should you use?

Results vary by use case, but every effort level lands higher on the quality curve than its Sonar counterpart — and in the deep-research tier, for less. Use the mapping below as a starting point, then confirm against your own traffic.
For state-of-the-art deep research, use the xhigh effort level — the highest quality in the lineup, leading across the benchmarks we run.

Explore the effort levels

See what each effort level configures and how to override individual settings.

Migrate your integration

Follow the field-by-field procedure to move a Sonar integration to the Agent API.