One dial in Perplexity Computer sets model and cost

The most useful line in this rollout is the one that names the plumbing: effort presets "combine the orchestrator model and reasoning depth to control how deeply and efficiently Computer works through a task." That is Perplexity's own description of the effort and speed selector now reaching Perplexity Computer, in a post on Threads.

At a glance

• Perplexity Computer is getting a single control that picks which model runs the orchestrator and how much reasoning it applies, instead of leaving you to assemble that pair by hand.
• According to SQ Magazine, lower settings use less expensive models, so what a user spends tracks both the setting picked and the work the agent actually performed.
• SQ Magazine has the selector live on the web only, with Android and iOS support coming and no date given, while custom model and reasoning settings stay for anyone who prefers them.

If you have not been keeping score on Computer, the agent has been assembled in public over a few months. According to TestingCatalog, Perplexity released Personal Computer to all Max subscribers on 17 April 2026, shipped the Brain memory system for Computer on 19 June 2026, and moved every Computer session onto a sandboxed runtime called SPACE on 17 July 2026. The effort dial lands on top of that stack.

Effort presets fold the orchestrator model and reasoning depth into one pick

The sentence rewards unpacking. The orchestrator is the model that plans the job and fires the tool calls; reasoning depth is how much thinking it is allowed to spend before it acts. A preset bundles those two decisions, so one pick sets both who plans and how long the planning runs. The feature's name pairs the same two ideas: effort and speed.

The hand-built route survives. SQ Magazine says custom controls remain for people who want to name a specific model and set its reasoning level themselves, which leaves the presets as the fast path rather than the only path.

What does the cheap end of the dial actually cost you?

Here is the part that touches the bill. According to SQ Magazine, lower settings use less expensive models, so what a user spends tracks the setting picked and the work performed; the same report has the controls live on the web only.

Perplexity has other cost levers in flight. TestingCatalog reported that the company was preparing a Hybrid mode for Computer on Mac that delegates some subtasks to locally hosted models to save costs, with a Privacy Gate evaluating data for personally identifiable information before it goes to the cloud, and that it was testing a way for subscribers to connect OpenRouter accounts, potentially unlocking cheaper models for Computer tasks.

The orchestrator plans the task and fires the tool calls

Think of it as a dispatcher rather than a driver. In agent systems generally, as JetBrains describes the pattern, the orchestrator breaks an objective into tasks, keeps them in a queue, checks each task's dependencies, routes it to the right agent or tool, and waits for a result before updating the workflow state; after every step it decides whether to proceed, retry, or escalate.

The scale around that dispatcher is what makes the choice of model non-trivial. Per Aibyaakash, Computer orchestrates 19 AI models, connects to more than 400 apps, and runs entire workflows in the cloud while the user's laptop is closed. JetBrains adds that scaling this pattern usually means distributing task coordination through message queues or event streams while workflow-level decisions stay centralized.

The inputs are thin on exactly the parts a paying user would check first. Nobody has published the preset names, a price for any step of the dial, or a figure for how much cheaper the low end works out to be, even though SQ Magazine ties the spend to the setting and the work done. Oddly, for a control sold on efficiency, efficiency is the one quantity left unstated.

When the dial reaches phones

Mobile is the next checkpoint, and it has no date: SQ Magazine says Android and iOS support is coming without one, so today this is a web feature. The open question is what sits behind each preset, since the list of orchestrator models and their reasoning levels has not been published. The Mac Hybrid mode and the OpenRouter connection that TestingCatalog described also remain unshipped.