Skip to main content

CORNERSTONE FAQ

What is Perplexity Computer and should I use it?

A multi-agent system. Opus 4.6 orchestrates, specialist models execute. We use it to run our shop. The pricing reality, the real use cases, and where it should not touch.

Reading time 6 min Updated May 2026 Topic Perplexity

The short answer

Perplexity Computer is a multi-agent AI system. One main reasoning model (Claude Opus 4.6 at the time of writing) acts as the orchestrator and assigns specialized sub-tasks to a roster of other models: Gemini Deep Research for long-context document review, Grok 4.1 Fast for cheap throughput, GPT-5.2 for long-context tasks, Sonar for retrieval, Nano Banana for images, Veo 3.1 for video. The user describes a goal. The system picks the model. We use it daily to run Restrepo Innovations. We pay full price at the Max tier ($200/month). We are not affiliates.

What it actually is

Strip the marketing away and Computer is a productized version of the router pattern. Instead of a single chatbot that does everything badly, you have an orchestrator that picks the right specialist and a set of tools the orchestrator can call: web browser, code interpreter, document writer, image generator, calendar reader, email drafter. The model decides what to call and in what order. The user reviews the result.

The model lineup under the hood

  • Orchestrator: Claude Opus 4.6 (the planner, plans the workflow).
  • Deep research: Gemini 3 Pro (long context, multimodal).
  • Throughput: Grok 4.1 Fast (cheap, fast, used for parallel low-stakes calls).
  • Long context: GPT-5.2 (used when the document set exceeds Gemini’s window).
  • Retrieval: Sonar / Sonar Pro (Perplexity’s own search).
  • Images: Nano Banana.
  • Video: Veo 3.1.

Composition changes as Perplexity adds and rotates models. The user does not have to track it.

Connectors and what they unlock

Computer connects to Gmail, Google Calendar, Google Drive, GitHub, Notion, Slack, Outlook, Microsoft 365, and an expanding list of others, with custom MCP (Model Context Protocol) connectors filling the gaps. The connectors are what make Computer different from a chat session. With connectors active, Computer can read your inbox, write to your Drive, post to your Slack, schedule on your calendar, and commit to your GitHub. With permission. Always with permission, and with confirmation on outbound actions that matter.

Real Restrepo use cases

  • Blog cron. Every Monday morning, Computer pulls the prior week’s Search Console queries, identifies a high-opportunity emerging keyword, drafts a post in Restrepo voice, generates a unique hero image, FTPs the file to staging, pings IndexNow, and posts a Slack summary. We review and publish.
  • Query mining. Computer monitors search trends in our service area and flags topics worth writing about before the volume peaks.
  • FTP and deployment automation. File handling, cache invalidation, IndexNow pings.
  • Email drafting. Inquiry responses drafted by Computer, reviewed and signed by us.
  • Document research. Long PDF packages summarized to one page with conflicts flagged.
  • Vendor and pricing research. Computer compares product options across distributors and surfaces relevant differences.

Pricing reality

Pro at $20/month is the entry tier and covers most personal research and writing. Max at $200/month is the Computer tier; this is what we pay. Enterprise pricing for teams runs in the hundreds to low thousands per month depending on seats and connector volume. Heavy multi-hour Computer workflows can drain the Max monthly allowance in two to three weeks of aggressive use. Track your usage if you plan to run Computer hard.

What it cannot do safely

  • Numerical work in spreadsheets. LLMs are confidently wrong about results. Verify every formula.
  • Legal or financial drafting. Use it as a polish step, not a draft step. The voice should be yours.
  • Anything irreversible. Sending a calendar invite is fine. Wiring money is not. Computer should ask for confirmation on anything you cannot undo.
  • High-stakes client emails. Same rule. We draft, Computer polishes, we sign.
  • Smart home control. Computer should not touch your Crestron processor or Lutron hub directly. Voice commands and home automation belong on a deterministic stack with bounded scope.

RISK

The most common Computer mistake we see is letting it auto-send. The model is right most of the time. The cost of the wrong send (a misquoted price, a wrong recipient, a hallucinated commitment) is not worth the saved click.

How a luxury homeowner could use it

  • Concierge: "Find a private chef with these credentials in Greenwich next Saturday and draft an inquiry."
  • Travel pre-flight: "Every PDF on this trip, summarized, on a single page."
  • Real estate watch: "Flag listings in Bergen County over $5M that match these criteria."
  • Document review: "Compare these three orthodontist proposals for the kids."
  • Inbox triage: "Summarize my inbox from yesterday and draft replies for the routine ones."

The privacy footnote (still cloud)

Computer is cloud. Your queries go to Perplexity’s servers. Pro and Max plans do not train on your data per Perplexity’s stated policy. None of it is local. For genuinely sensitive work, run a local LLM in parallel and route the sensitive subset of work to it. We configure that hybrid. See Build It Yourself.

Should you use it?

If you spend two or more hours a week on tasks Computer can plausibly handle, yes. If you have sensitive operational data you would not want logged on someone else’s server, use it for the non-sensitive subset and keep a local stack for the rest. If you want help configuring it for your business or household, that is exactly the kind of project we take. Call 201.405.2022.

Related

Want AI in your home or building, done honestly?

We design and deploy AI that lives on hardware you own. No cloud lock-in, no data trades, no surprises. Serving NJ, NY, and CT metro.