Perplexity Comet & the AI Personal Computer: Everything Explained

The Full Context: What Perplexity Is Building in 2026

Perplexity started as an answer engine. That framing — a better search experience with citations — was accurate through most of 2024. By early 2026, it’s substantially incomplete.

What Perplexity is actually building is a three-layer AI operating environment: a cloud research and orchestration platform (Perplexity Computer), a local persistent agent (Personal Computer), and an AI-native browser to connect them to the web (Comet). Each layer is independently useful. Together they’re Perplexity’s answer to the question nobody in the industry can yet answer cleanly: what does it look like when AI stops being a chatbot and becomes infrastructure?

Three things happened in quick succession that shifted the company’s product trajectory. First, Perplexity removed advertising from its answers in February 2026 — a subscription-first pivot that signals a genuine commitment to answer quality over reach. Second, it launched Model Council the same month, treating multi-model orchestration as a product feature rather than back-end infrastructure. Third, it announced Personal Computer at Ask 2026 in March — a product built around a $599 Mac mini that it positioned not as a gadget but as infrastructure.

Samsung pre-installing Perplexity on the Galaxy S26 — the first non-Google company to get OS-level access — added distribution that no browser launch can manufacture. That partnership matters because it gives Perplexity ambient presence on the device most people use most of the time.

That sentence is clever marketing, and it’s also genuinely true. The Mac mini-as-AI-server framing is the clearest articulation I’ve seen of what “personal AI” actually means at the infrastructure level. You’re not renting cloud time. You’re running a persistent process on hardware you own, in your home or office, that happens to be orchestrated by Perplexity’s models.

Whether that’s better or worse than pure-cloud alternatives depends almost entirely on your specific workflow. I’ll get to that in detail.

Key Timeline: Perplexity’s 2025–2026 Product Arc

Comet Browser — Feature-by-Feature Breakdown

Comet is built on Chromium, which means Chrome extensions work, bookmarks transfer, and the rendering engine is familiar. That was a deliberate choice after Perplexity’s much-publicized bid to acquire Google’s Chrome browser (reportedly valued at $34.5 billion) fell through. Using an open-source foundation means they could ship fast. It also means the browser’s differentiation has to come entirely from the AI layer on top — because underneath, it’s structurally the same as Chrome.

That AI layer is genuinely different in architecture from what exists in Chrome extensions or Edge Copilot. The difference isn’t cosmetic. It’s about access depth.

Core Capabilities

Real-World Testing: 14 Days Replacing Chrome

I ran Comet as my primary browser on a MacBook Pro for two weeks — no Chrome fallback allowed except for specific sites that threw compatibility errors. Here’s what I found, without softening the failures.

Where Comet Genuinely Struggles

The memory problem deserves more explanation than most reviews give it. Comet’s context is session-scoped by default. Once you close a workspace, it doesn’t reliably remember what happened in it. For a product pitching itself as your personal AI assistant, this is a significant architectural gap — not a bug, but a deliberate limitation that Perplexity hasn’t fully addressed yet.

The agentic shopping conflict with Amazon (more on this in the legal section) also constrains one of Comet’s headline use cases. If you were planning to use it as an autonomous shopper across major e-commerce platforms, that use case is currently litigated and restricted on at least one major platform.

And the browser market itself is fiercely competitive in 2026. OpenAI’s Atlas browser integrates deeply with ChatGPT. The Browser Company’s Dia is more polished for workflow automation on Mac. Comet has the broadest platform availability (iOS, Android, Windows, Mac) and the most ambitious agentic architecture — but “most ambitious” and “most reliable” are not synonyms at this stage.

Personal Computer — The Always-On AI Agent

Personal Computer is conceptually bolder than Comet and operationally more specific. It’s not a browser. It’s a persistent AI process running on your local hardware, integrated with your file system and native Mac applications, that can work autonomously while you’re doing other things — or sleeping.

The pitch: wake up to a completed research brief, a drafted batch of emails, a compiled morning digest, a flagged set of documents organized by project. The AI worked while you didn’t.

The reality in May 2026: it mostly delivers on this for well-defined, repeatable tasks. It struggles with anything requiring judgment calls, and the credit system on Pro limits how much it can run before you hit a wall.

How It Works Technically

Double-press both Command keys on your Mac to summon it. Accept text or voice input. The agent then coordinates across Perplexity’s 19-model cloud infrastructure to plan and execute the task — using Claude Opus 4.6 as the primary orchestrator and GPT-5.5 as a secondary default (as of May 2026’s update). It can:

• Access your local file system (with explicit permission for each folder)
• Control any Mac app — Messages, Mail, Calendar, Notes, Pages, Numbers, Safari, Terminal
• Initiate and monitor Comet browser sessions for web-based sub-tasks
• Receive remote instructions from your iPhone
• Run on a schedule: “Every weekday at 6:45am, compile overnight news relevant to my three open projects and put a summary in /Documents/Briefs”
• Execute multi-step workflows with full audit trails

Every action that modifies data requires explicit approval. There’s a kill switch that halts everything instantly. The audit trail is comprehensive — you can see exactly what the agent did, in what order, and with what tools.

Real Use Cases and Workflow Patterns

The most credible use cases I’ve encountered from actual users fall into three categories. I’ll be specific rather than vague, because vague claims are how AI demos lie.

Research-Heavy Work (Works Well)

A freelance analyst running deep-dive reports describes it as running a Mac mini as “a dedicated research overnight worker.” She sets tasks before she finishes her day — pull recent filings from these 6 companies, cross-reference with the quarterly notes in this folder, draft a comparison table, save it in the project directory. She opens her laptop in the morning and the draft is there. This is the scenario Personal Computer was built for, and it delivers it reliably enough to be genuinely useful.

File Organization and Email Triage (Works With Caveats)

The email triage use case (draft responses, categorize incoming, flag priority items) works for straightforward correspondence. It breaks down on anything requiring nuanced judgment — knowing that a “quick question” from a client might actually be a contract renegotiation in disguise isn’t something the agent handles gracefully. Use it to draft, not to decide.

Coding and Deployment Tasks (Promising, Early)

Via Perplexity Computer in Microsoft Teams (an integration launched in May 2026), engineering teams can run agentic coding tasks at the team level. Running test suites, compiling reports, generating documentation drafts — these work. Anything touching production infrastructure should wait for a more mature release.

The Mac Mini Question

Perplexity recommends the Mac mini specifically — not as a sponsorship play but for practical reasons. It’s cheap relative to other Apple hardware ($599 starting price), silent, designed to run unattended, consumes very little power in standby, and Apple’s M-series chips handle local context operations efficiently. You can leave it plugged in 24/7 without meaningful noise, heat, or electricity cost concerns.

Tim Cook mentioned on Apple’s Q2 2026 earnings call that Mac mini and Mac Studio were supply-constrained “for months to come” — a direct consequence of AI platforms like Perplexity driving unanticipated demand.

Model Council and the 19-Model Architecture

Model Council launched February 5, 2026, and it’s one of the more intellectually honest product decisions Perplexity has made. The premise is straightforward: different AI models have different blind spots, and nobody reliably knows which model is best for a given query type. Rather than forcing users to guess or run comparisons manually across tabs, Model Council automates that triangulation.

Submit one query. Three frontier models — Claude Opus 4.6, GPT-5.2, Gemini 3.0 — process it simultaneously. A fourth synthesizer model resolves conflicts, surfaces points of agreement, and presents one consolidated response while flagging where models diverged significantly.

The broader Perplexity Computer architecture runs 19 specialized models. Not all of them are frontier models — the total includes task-specific sub-agents for things like web navigation, code execution, file operations, and calendar management. The orchestration layer routes sub-tasks to the appropriate specialist rather than running everything through one general model.

This is the most interesting technical bet Perplexity has made. The assumption is that model specialization will continue to widen — that the performance gap between “best general model” and “best model for this specific task type” will grow, not shrink. If that’s correct, orchestration infrastructure becomes the durable competitive advantage, not any individual model.

Pricing Reality Check: What You’re Actually Paying For

Pricing has shifted several times in 2026. Here’s the current state as of May 25, 2026:

Also: Comet Plus at $5/month unlocks premium publisher content within Comet’s reading experience. It’s available as an add-on to any main plan and is separate from the AI feature tiers.

Browser Wars: Comet vs. Atlas vs. Dia

There are three serious AI-native browsers in 2026. All three are Chromium-based. All three have meaningfully different philosophies. Here’s an honest breakdown:

The Amazon Legal Battle and What It Signals

This is the most important context most coverage of Comet underplays. In November 2025, Amazon sued Perplexity in San Francisco federal court. The core allegation: Comet’s shopping agent accessed password-protected Amazon user accounts while disguising itself as a standard Chrome browser — not an AI agent — to bypass Amazon’s bot-detection systems. Amazon argued this constituted unauthorized access under the Computer Fraud and Abuse Act (CFAA), a 1986 law that predates the web by several years.

On March 10, 2026, District Judge Maxine Chesney granted Amazon a preliminary injunction. The ruling drew a distinction that will shape the entire agentic AI industry: user authorization and platform authorization are separate legal requirements. A user giving Comet their Amazon login credentials did not automatically give Comet the legal right to use those credentials on Amazon’s platform. Judge Chesney found Amazon had provided “strong evidence” that Perplexity accessed accounts “with the Amazon user’s permission but without authorization by Amazon.”

Perplexity appealed immediately. The 9th Circuit temporarily stayed the injunction while considering the appeal. The EFF, Mozilla, and other digital rights groups filed amicus briefs in April arguing the ruling represents an “alarming expansion” of the CFAA that would fundamentally undermine the concept of agents acting on users’ behalf online.

The industry implications extend beyond Perplexity. If the injunction stands or the underlying legal principle holds, every AI agent that takes actions on behalf of users on third-party platforms faces the same legal exposure. The model — user says “buy this for me,” agent goes and buys it — requires platform authorization, not just user consent. That changes the entire architecture of agentic commerce.

Privacy and Security: The Honest Assessment

There are several distinct privacy questions here that tend to get conflated, so let’s separate them.

Comet Browser Privacy

Comet uses Perplexity’s AI search engine as its core. Perplexity’s business model is subscriptions now (ads removed February 2026), which removes the incentive to monetize query data through advertising. But queries still pass through Perplexity’s servers. The assistant, when given access to your email or calendar via connectors, reads that data to improve context. Whether that data is retained, for how long, and how it’s used in model training is covered in the privacy policy — which most users don’t read and which has changed multiple times over the past 18 months.

Personal Computer Privacy

This is where the gap between marketing and reality is most significant. Personal Computer is positioned around local operation — your Mac, your files, your apps. But it’s not a locally-run AI. The orchestration happens on Perplexity’s cloud. When you ask it to “review this confidential client contract and draft a response,” the content of that document passes through Perplexity’s infrastructure. For anything covered by NDA, client confidentiality agreements, healthcare privacy rules (HIPAA), or financial data regulations, this matters considerably.

The positive side: Perplexity’s explicit approval model for every sensitive action, combined with a full audit trail, is better than most AI agent products. You can see exactly what the agent did. The kill switch works immediately. These aren’t cosmetic trust features — they’re functional controls. But they don’t address the cloud-routing question for sensitive data.

Who Should Use This (And Who Definitely Shouldn’t)

Use Comet if:

• You spend 3+ hours daily in a browser doing research, reading, and synthesizing information across multiple sources
• You want cross-platform mobile continuity — this is the only serious AI browser on both iOS and Android as of May 2026
• You’re a student, journalist, analyst, or researcher who treats the browser as a primary work environment
• You want free agentic features without a paywall — Comet’s agent mode is free, which is genuinely unusual
• You’re already in the Perplexity ecosystem and want tighter integration

Don’t use Comet as your only browser if:

• You rely on a specific extension that has known Chromium compatibility issues
• You do heavy agentic shopping and e-commerce tasks — the Amazon situation is unresolved
• Battery life is critical and you’re frequently off-power
• You need reliable cross-session memory for ongoing research threads

Use Personal Computer if:

• You have a Mac running Sonoma 14+ and can dedicate a machine (ideally Mac mini) to always-on operation
• You run research-heavy, file-intensive workflows where overnight autonomous operation has clear ROI
• You’re comfortable with the cloud orchestration model and your workflows don’t involve regulated or highly confidential data
• You’re on Pro ($17–20/mo) and want to evaluate before committing to Max — now possible as of May 7

Don’t use Personal Computer if:

• You’re on Windows — there is no timeline for Windows release
• Your workflows involve confidential client data, HIPAA-covered information, or anything regulated
• You need reliable cross-session memory — it’s still the biggest v1.0 gap
• You expect it to replace a human assistant for judgment-intensive tasks — it doesn’t yet
• You’re evaluating it for enterprise deployment without legal review of the Amazon litigation context

Final Verdict

Perplexity shipped more genuinely interesting product decisions in the first five months of 2026 than most AI companies managed in all of 2025. The subscription pivot is credible. The 19-model architecture is technically ambitious in the right direction. Model Council is an honest solution to a real problem. Personal Computer’s local-plus-cloud architecture is the right mental model for what personal AI should look like.

And yet.

Comet’s memory is still session-scoped in a product that markets persistent intelligence. Personal Computer’s “local” framing elides the cloud routing of sensitive data. The Amazon litigation is unresolved and the legal exposure for agentic commerce is real. Windows users don’t have a timeline. Credit limits on Pro bite on the workflows where the product is most compelling.

These aren’t fatal flaws. They’re the expected rough edges of a v1.0 product family pushing into genuinely new territory. The question is whether you’re the right person to be working through those edges right now.

If your work is research-heavy, Mac-based, and you can afford the credits, Personal Computer is worth serious evaluation. If you read a lot and want an AI that actually lives inside your browser rather than bolted onto it, Comet is worth switching to for a two-week trial — the download is free, the cost of trying is time and the mild inconvenience of resetting your defaults afterward.

For everyone else: Perplexity free tier plus good prompting discipline gets you surprisingly far, and the products will be meaningfully better in six months.