Why this page exists
If a vendor is going to recommend a tool, the vendor should be using the tool. We use Perplexity. We pay full price for it. We are not affiliates. We have no commercial relationship with Perplexity. The blog post you might be reading next on this site was queued by Perplexity Computer; the hero image was generated through it; the publishing pipeline runs on a Monday cron it manages. That is the disclosure. The rest of this page is the working review.
What Perplexity is
Perplexity is an answer engine. It was founded in 2022 by Aravind Srinivas (formerly OpenAI, DeepMind), Denis Yarats, Johnny Ho, and Andy Konwinski. As of mid-2026 the company is valued at roughly $21B and serves about 45 million monthly users. Its product is a search box that returns a written answer with citations, instead of a list of blue links. Underneath, it routes queries to a mix of frontier models (Claude, GPT, Gemini, Grok) and its own Sonar model family.
The product line in plain English
- Perplexity (the answer engine). Free at perplexity.ai. The default product. Search box, written answer, citations.
- Sonar and Sonar Pro API. Developer-grade access for building Perplexity into your own apps. We use this for internal tooling.
- Comet. The Perplexity AI browser. A browser that browses for you: takes natural-language instructions, opens tabs, fills forms, scrapes data, summarizes findings.
- Perplexity Computer. The multi-agent system. Opus 4.6 acts as orchestrator, farms tasks to specialist sub-agents, runs workflows for hours.
- Pricing tiers. Pro at $20/month. Max at $200/month (the Computer tier). Enterprise pricing on request.
What Perplexity does well
Citations on every claim
This is the feature that distinguishes Perplexity from ChatGPT, Gemini, and Claude on default behavior. Every assertion in a Perplexity answer ships with at least one source link. You can verify the answer in the time it takes to click. ChatGPT now offers similar in some modes; Perplexity made it the default from day one and the implementation is more consistent.
Live web access without prompting
Perplexity hits the live web by default. ChatGPT requires you to engage browse mode. Claude does not browse without a tool call. Gemini browses on certain queries. Perplexity always does. For research-style questions where freshness matters, this is the right default.
Source diversity
Perplexity reports citing 1,430 unique news sources, compared to roughly 881 for Google and 707 for OpenAI’s search. Whether or not you trust those exact numbers, the qualitative experience matches: Perplexity surfaces smaller publications more often, which is useful for both research breadth and source triangulation.
Subscription-first, not ad-funded
Perplexity briefly experimented with ads in 2024 and 2025 and pivoted away in February 2026 toward a subscription-first model. That alignment matters. An ad-funded answer engine has structural pressure to surface the answer that pays best, not the answer that is correct. A subscription-funded answer engine has structural pressure to be useful enough that you keep paying.
What it does not do well
The CJR 37% audit
A Columbia Journalism Review audit of Perplexity answers found 37% contained errors. Citations are not accuracy. The model can cite a source and still misrepresent what the source says. We have seen this in our own usage. The defense is to actually click the citation when the stakes are above casual.
Pro Search overconfidence
Perplexity Pro Search (the deeper-research mode) is sometimes more confidently wrong than the free version. The deeper search retrieves more sources, which can make the resulting answer feel more authoritative even when the synthesis is weaker. We use Pro Search for breadth, then verify before publishing.
Not a primary-source verifier
For legal, medical, or financial work, Perplexity is a starting point, never an endpoint. You do not file a brief based on a Perplexity citation any more than you would have filed one based on a Wikipedia citation in 2010. We treat it as a high-quality research assistant, not as a court-admissible source.
Comet at its worst
Comet is impressive when it works. When it fails, it can be slower than typing the URL yourself. The browser sometimes loops on form fields, sometimes misidentifies UI elements, sometimes refuses to authenticate to sites it has authenticated to before. The trajectory is upward; the current state requires patience.
Perplexity Computer specifically
Perplexity Computer is the headline product. Architecturally, it is a multi-agent orchestrator. One main reasoning model (Claude Opus 4.6 at the time of writing) sits as the planner and coordinator. It farms specialized sub-tasks to a roster of specialist models:
- Gemini Deep Research for long-context document and web review
- Nano Banana for image generation
- Veo 3.1 for video
- Grok 4.1 Fast for high-throughput cheap queries
- GPT-5.2 for long-context tasks where its window helps
- Sonar for retrieval
The orchestrator picks. The user does not. From the operator’s seat, you describe the goal and the system picks the model. This is the router pattern productized.
Connectors
Computer connects to Gmail, Google Calendar, Google Drive, GitHub, Notion, Slack, Outlook, Microsoft 365, and a growing list of others. It can read and write through these connectors with documented permission scopes. Custom connectors via MCP (Model Context Protocol) are increasingly common.
What it can do
It can run a browser tab, write code, build PDFs and Word documents, send email (with confirmation), schedule events, browse the web for hours on a single instruction, monitor a folder for new files and act on them. It runs in the cloud, so you can close your laptop and the workflow continues.
A real Restrepo workflow
Our Monday morning blog cron is illustrative. Computer wakes up at 6 a.m. ET. It pulls the prior week’s Search Console queries via the connector, identifies the highest-opportunity emerging keyword, drafts a post in Restrepo voice using a style guide stored in Drive, generates a unique hero image through Nano Banana, lays out the HTML using our internal template, FTPs the file to our staging server, pings IndexNow on the published URL, and posts a Slack message in the team channel summarizing what it did. Total human time: about ten minutes of review before the Slack post. Total human time without Computer: roughly four to six hours.
The honest cost picture
Pro at $20 per month is the entry tier and covers most individual research and writing usage. Max at $200 per 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 seat count and connector volume.
The credit economics on Computer specifically: heavy multi-hour workflows can drain a Max account’s monthly allowance in two or three weeks of aggressive use. We track our usage. For our shop, the math comes out clearly favorable: $200 a month replaces work that would otherwise cost a part-time associate. For a luxury homeowner using Computer for personal concierge tasks, the Pro tier is usually plenty.
Where we still do not trust it
- Numerical work in spreadsheets. We verify every formula. LLMs are fluent in Excel syntax and confidently wrong about results.
- Anything that touches legal or financial wording. We draft. We have it polish. Never the reverse.
- High-stakes client emails. Same rule. The voice is ours. The polish is the model’s.
- Anything irreversible. Sending a calendar invite is fine. Wiring money is not. The line is between actions you can undo and actions you cannot.
How a luxury homeowner could use Perplexity Computer
- Concierge. "Find a private chef with these credentials available in Greenwich next Saturday, draft the inquiry, hold for my review." Computer handles the search, draft, and queue. Owner clicks send.
- Travel pre-flight. "Every PDF on this trip, summarized, on a single page with the conflicts highlighted." Connector reads Drive, model summarizes, output is one page.
- Real estate watch. "Watch listings in Bergen County over $5M and alert me when something hits these criteria." Cron job, daily check, email summary.
- Family operations. "Compare these three orthodontist proposals for the kids and rank them." Document review and draft summary.
Where it should not touch: primary financial moves, anything irreversible without confirmation, anything safety-relevant in the home. Computer should never speak to your Crestron processor without a human in the loop.
The privacy footnote: Perplexity is still cloud
Your queries go to Perplexity’s servers. Pro and Max plans do not train on your data per Perplexity’s stated policy. Free tier may. None of it is local. For private operations -- family financial information, sensitive client data, anything you would not want logged on someone else’s server -- we recommend running a local LLM in parallel and routing the sensitive subset of your work through it. See Build It Yourself for that architecture.
Restrepo + Perplexity: how we will help
For business clients, we configure Computer with the right business connectors (Drive, Gmail, Notion, Slack, custom MCP if needed) and build the workflows that actually save you time. For luxury homeowners, we deploy a hybrid: Pro or Max for non-sensitive research and concierge work, plus a local LLM (Llama on a Mac Studio) for anything that should never leave the property. We train your team on the router-pattern workflow so they know which queries belong on which tier.
Comet specifically: the AI browser
Comet is a Chromium-based browser with a Perplexity assistant baked in. The pitch: instead of opening five tabs, summarizing them mentally, and synthesizing an answer, you ask Comet to do it and the browser opens the tabs, reads them, and produces the answer. The promise is real for research-style tasks. The reality is that Comet is fastest when you are doing something the browser has clearly been trained on (review aggregation, comparison shopping, news summarization) and slowest when you are doing something it has not (logging into a site with multi-factor, navigating a niche admin panel, anything where the page layout is unusual).
Our usage pattern: Comet for breadth-first research where we want a synthesized answer across many pages. Standard browser for depth-first work where we know exactly what we are looking for. We do not use Comet for any session that includes financial logins. The trust model for an AI browser that can act on your behalf is still maturing, and the prudent posture is to limit what it can touch.
How Perplexity differs from ChatGPT, Claude, and Gemini
The frontier chatbots and Perplexity overlap in capability but differ in default behavior. ChatGPT, Claude, and Gemini default to drawing on training data with optional browse modes. Perplexity defaults to browsing with optional reasoning. For factual questions where freshness matters, Perplexity wins. For long writing or code where reasoning matters more than freshness, Claude or GPT win. For broad research with the largest context window, Gemini wins. The router pattern lets you stop choosing.
Perplexity Computer collapses the choice for you. The orchestrator picks the right model for each sub-task. From the user’s seat, it looks like one assistant. Underneath, it is the same router pattern we have always recommended, just productized. We pay for the productization because it works.
Real numbers from our shop
Approximate, May 2026, on a single Max account ($200/month):
- Around 12 to 18 hours of Computer workflow time per week.
- Around 8 to 12 long-form research sessions on the Pro tier per week (separate account, additional $20).
- Around 30 to 50 short queries per day across the team on Pro tier.
- Approximate replaced labor hours: 15 to 25 hours per week of routine work that used to consume associate time.
- Approximate cost replaced: at modest associate billing rates, $1,500 to $3,000 per week of work output, against $220 per month in tooling.
The math is favorable. The math would not be favorable if we did not also have a human review step on every output that goes outside the company. The unreviewed savings would be smaller and would carry more risk than the reviewed savings carry value.
The bottom line
Perplexity is the best general-purpose AI tool we have used. It is also not magic, not free of error, and not local. Use it for what it is good at, verify what matters, and keep a private fallback for what does not belong on a cloud server. That is the working integrator’s position.