Computer adds effort controls for model selection

Today, we’re rolling out effort controls in Perplexity Computer, giving people more control over how much effort it puts into their work. Using a slider, they can choose one of four settings, from Light to Ultra, and Perplexity selects the model and reasoning level for each setting.

People told us they wanted to decide how much effort a task deserved and what they were willing to spend, while leaving model selection to Perplexity. Users who prefer to choose a specific model and reasoning level can still do so in custom controls.

Four settings for different kinds of work

Users select the setting they want to use from the slider in the omnibar:

• Light: For straightforward, everyday tasks, such as collecting vendor invoices in a spreadsheet.
• Standard: For balanced reasoning and cost on everyday work, such as combining account notes and company updates into a sales brief.
• High: For complex analysis, such as weighing customer requests against engineering estimates to prioritize product work.
• Ultra: For maximum effort on open-ended problems, such as evaluating a new market under different growth assumptions.

Lower settings use less expensive models. Credit usage depends on the work Computer performs.

Expand

More effort for demanding work

An effort setting selects the model responsible for coordinating an assignment and how deeply it reasons. The orchestrator decides how to approach the work and delegates parts of it to supporting agents, which can use models from different providers.

One example: A memo informing an important decision may warrant deeper analysis before someone acts on its recommendation. Users can choose a higher setting for that assignment without having to choose a specific model or configuring its reasoning level themselves.

Model selection informed by billions of queries

Perplexity’s experience answering billions of queries informs how we match models to work. As people use Perplexity for a wider range of assignments, we learn more about where different models perform well and where additional reasoning improves results.

We test models from different providers at different reasoning levels, comparing the quality of their work with its cost. When a lower-cost combination produces comparable results, users can keep more of their budget for assignments that benefit from deeper reasoning.

Access effort controls in the omnibar

The effort controls area available on web and will be available on Android and iOS soon. Users can access the effort slider in Computer, with Light, Standard, High, and Ultra settings.

Custom controls are still available for users who want to choose a specific model and reasoning level.