Perplexity Computer: Automating Multi-Step Projects With 20+ AI Models (2026)

Most AI assistants answer one question at a time. Perplexity Computer takes a single complex prompt, breaks it into subtasks, and coordinates a team of specialized AI agents to carry the whole project from start to finish. Launched in February 2026, it is a meta-orchestration system: it runs inside an isolated compute environment with a real filesystem, a browser, and tool integrations, and it can keep working for hours.

The shift is from a chat interface to a delegated workspace. You describe an outcome, such as a researched report, a deployable website, or a populated spreadsheet, and Computer plans the steps, assigns each one to the model best suited for it, runs them asynchronously, and assembles the result. A live credit counter shows the cost accruing in real time as the work proceeds.

This breakdown explains what Perplexity Computer is, how its multi-model orchestration works, the tiers and pricing that gate access, the connectors and workflows it supports, real examples from independent reviewers, and the limitations those reviewers flagged. Pricing, credit limits, and the specific models in the orchestrator move quickly, so treat the figures here as a snapshot of early-to-mid 2026 and confirm current details with Perplexity before you commit.

What Is Perplexity Computer?

Perplexity Computer is a meta-orchestration system. Rather than a single model answering a single prompt, it takes one complex instruction, decomposes it into subtasks, and asynchronously coordinates specialized sub-agents to complete the work end-to-end. The whole thing runs in an isolated compute environment that includes a real filesystem, a browser, and connections to external tools, which is what lets it act on the world rather than just describe it.

That environment is the difference between a chatbot and an agent. A chatbot returns text. Computer can open a browser, navigate to a site, read a file you created, write a new file, query a database, and deploy a website, then hand back the finished artifact. Because the tasks run asynchronously and the environment persists, a single job can run for hours without you watching it.

It launched in February 2026 and expanded quickly. The timeline below traces the rollout from the initial Max-tier launch through the March additions for Pro and Enterprise.

For more context on where this sits in the broader ecosystem, see the AI tools hub and the Perplexity sub-hub.

How the Orchestration Works

The defining idea behind Computer is that no single model is best at everything, so it routes each subtask to the model suited for it and runs them in parallel. At the February launch it coordinated roughly 19 specialized models; by March the count had grown to 20 or more. The exact lineup shifts release to release, so treat specific model names as a snapshot of early-to-mid 2026 rather than a fixed roster.

As of that window, the changelog described a division of labor along these lines:

• Claude Opus handled core reasoning and planning, the work of breaking a prompt into a coherent sequence of steps.
• Gemini 3.1 Pro drove deep research, the long-horizon retrieval and synthesis tasks.
• Grok covered lightweight, speed-sensitive steps where latency mattered more than depth.
• GPT-5.3-Codex took software engineering, the code generation and editing subtasks.

You are not locked into the automatic routing. Users can override the model choice for a given task, and two related features sit on top of the router. Model Council runs a prompt through several frontier models and synthesizes their outputs into one answer. Custom Skills let you save a reusable procedure once and invoke it across future jobs, so a workflow you tune for one project becomes a building block for the next.

Tiers & Access

Computer launched on the highest consumer tier and worked its way down. Access, credit allowances, and spend limits differ by plan, and because agentic tasks burn credits in real time, the credit math matters as much as the headline price.

Credit allowances and pricing move quickly. Confirm current credit limits and prices on Perplexity's site before subscribing. Figures reflect the February to March 2026 changelog.

Connectors & Workflows

An agent that can only read the open web is limited. Computer's reach comes from connectors: it ships with more than 400 prebuilt integrations, including Salesforce, HubSpot, Box, and Vercel, so it can pull from and act on the systems a business already runs. For data warehouses, a direct Snowflake connection generates an auto-built Data Map that translates natural language into SQL, so a plain-English question becomes a query against live tables.

If a system you use is not on the prebuilt list, the Bring Your Own Connector path uses the open Model Context Protocol (MCP) with OAuth or an API key, which means teams can wire in internal tools without waiting for an official integration.

Enterprise Workflows

Enterprise accounts get six packaged workflows, two of them new in this rollout. The two additions are website generation and website audit, the latter covering SEO and GEO, accessibility, and brand positioning. The rest span coding, design, and deployment, so a single job can move from building something to checking it to shipping it.

File Creation and Editing

Computer produces artifacts, not just chat replies. It drafts and formats documents, slide decks, spreadsheets, dashboards, and deployable websites. Editing is direct: you draw a selection box around part of an output and instruct the change in place. Drafting is Markdown-first, with export to PDF or DOCX on demand. A live credit counter tracks spend as the work runs, and scheduled task management lets jobs run on a recurring cadence.

Where It Runs

Computer reaches beyond the web app. A Slack integration responds to direct messages and channel mentions and can schedule recurring workflows. The Comet AI browser hosts agentic browsing, with a Comet Enterprise variant deployable through mobile device management (MDM). On Mac, a Personal Computer mode adds local file editing and local browsing, so the agent can work against files on your own machine rather than only in the cloud sandbox.

Real Examples

The clearest sense of what Computer does comes from watching it run a whole job. Three examples, one from an independent reviewer and two from the enterprise workflows, show the range.

Building an App in Minutes (Independent Review)

In a hands-on test, the team at Zapier asked Computer to build a meditation app. It ran the design, development, and audio subtasks in parallel and deployed a working web app it named "Serenity" in roughly eight minutes. The name and the app were Computer's output during that review, not a Perplexity marketing claim, and the audio portion was where the reviewer drew a line, as the limitations section notes.

Querying a Warehouse in Plain English

On the enterprise side, the Snowflake Data Map turns a business question into SQL. A representative prompt: "Pull quarterly revenue by vertical from Snowflake and compare year-over-year growth in a table." Computer translates that into the underlying query, runs it against live data, and returns a formatted table, no hand-written SQL required.

Scheduling a Recurring Report

Scheduled tasks plus the Slack integration cover routine reporting. A prompt such as "Every Monday at 9am, pull pipeline numbers from Salesforce and post a summary to #sales-updates" sets up a recurring workflow that runs on its own and posts the result to the channel, combining a connector, a schedule, and a delivery destination in one instruction.

Limitations to Know Before You Start

Perplexity, Perplexity Computer, Perplexity Max, Perplexity Pro, and Comet are trademarks of Perplexity AI, Inc. All other product names, logos, and brands are property of their respective owners. This article is editorially independent and not affiliated with or endorsed by Perplexity AI, Inc.