Perplexity

Perplexity in August 2026 is no longer just an answer engine with citations. The product now has several distinct layers: ordinary Search, Advanced Deep Research, a multi-model Search selector, Computer for multi-step cloud work, Personal Computer for local Mac access, Comet as an AI-native browser and Portable Computer as a newer local-first agent architecture. Reviewing “Perplexity” as though all of those surfaces were one model would therefore be misleading.¹²³

The core proposition remains research-first. Search finds and synthesizes web information with citations; Deep Research performs longer source collection, cross-checking, document work and code-assisted analysis; Computer moves from answering to acting across tools and services. Perplexity's own documentation describes the distinction simply: ordinary Perplexity gives answers, while Computer carries out work.⁴²

The current product is also explicitly multi-model. In Search, Pro and Max users can choose among Perplexity's Sonar 2 and third-party models including GPT-5.6 Terra, Gemini 3.1 Pro, Claude Sonnet 5, Grok 4.5, Kimi K3, GLM 5.2 and Nemotron 3 Ultra; Max additionally exposes higher-end models such as GPT-5.6 Sol and Claude Opus 5. Perplexity says the model selector is the source of truth because availability changes over time.¹

That model flexibility is one of Perplexity's strongest differentiators. It is also why a benchmark of one Sonar generation cannot stand in for the whole product.

Perplexity's core Search remains the easiest part of the product to understand. The user asks a current-information question; Perplexity retrieves web material, synthesizes an answer and exposes citations. The product can automatically choose a model or let eligible paid users select one manually.¹

That architecture has two advantages. First, source visibility makes verification easier than in a pure closed-book chatbot. Second, model choice lets the product behave more like a research interface over several frontier providers than a wrapper around one fixed model.

Tom's Guide's longer-term use found Perplexity particularly strong as a research assistant, comparison engine and productivity shortcut, praising synthesized answers, Deep Research, Spaces and model switching while still warning about hallucinations.¹⁶ TrustRadius's current synthesis likewise says verified reviewers consistently value research and analysis speed.⁸

The limitation is important: citations are not proof that every claim is faithful to the cited source. WIRED's older controlled test found a serious failure mode in which Perplexity produced an unsupported summary around a page it did not appear to have properly fetched. That investigation predates major product revisions, so AiToolMap does not treat it as a current 2026 quality score; it remains a historical warning about why source-grounded answers still need source checking.¹⁷

The current panel does not contain a sufficiently clean 2026 controlled factuality study that would justify declaring Perplexity more accurate than every rival. The right conclusion is narrower: its interface is unusually verification-friendly, but verification is still required.

Advanced Deep Research is designed for larger research jobs: it searches more sources, cross-references information, can analyze uploaded documents, run calculations in a code sandbox and produce editable reports. Perplexity also added clarifying questions, progress visibility and follow-up questions during an active run.⁴

Unlike ordinary Search, Deep Research chooses models automatically. Perplexity's current model-selection documentation explicitly separates the two modes: Search exposes eligible model choice, while Deep Research handles routing itself.¹

There is a small documentation-freshness complication worth preserving. A July Help Center page names specific Claude generations for Max and Pro Deep Research, while the general model-selector page says models change frequently and the current account selector/routing is authoritative. AiToolMap therefore treats the exact Deep Research backend as a dynamic implementation detail rather than a permanent plan promise.⁴¹

The product-level value proposition is stronger than the model label: more sources, iterative research, document ingestion and code-assisted analysis. Current first-party benchmark claims are favorable, but they remain vendor claims and are not treated as independent evidence.⁴

Computer is the largest conceptual change in the product. Perplexity describes it as a digital worker that can research, create documents and presentations, build applications, manage email, write code, work across connectors and continue scheduled or condition-based tasks.²

The architecture is explicitly agentic. Computer can create sub-agents for research, financial analysis, news monitoring and coordination; it can use Gmail, Outlook, GitHub, Linear, Slack, Notion, Snowflake, Databricks and Salesforce; and it runs in a personal cloud sandbox with persistent context.²

TechCrunch and VentureBeat both reported the launch as a multi-model orchestration bet rather than a single-model assistant. VentureBeat described the initial system as coordinating 19 models, while Perplexity's current Personal Computer documentation now says agent teams can span more than 20 frontier models.¹⁸¹⁹³

DataCamp provides a useful independent hands-on complement. Its May 2026 test ran a fixed parallel research workflow and examined how sub-agents, planning and visible credits behaved in practice. The reviewer specifically notes that launch-era descriptions became stale quickly as model routing and credit controls changed.¹³ That is precisely why this review relies on current first-party credit documentation for the economics rather than freezing May settings as permanent.

Personal Computer is the Mac app rebuilt around Computer. It can read and write local files, control native applications such as Mail, Finder, Slack, Messages and Notes, accept voice commands and remain always on. A Mac mini is recommended for continuous background use but is not required; any Mac on macOS 14 Sonoma or later is supported.³

Perplexity says sensitive actions are protected by on-device authorization and execution occurs in an isolated sandbox; Mac mini sessions use two-factor authentication. The app can also be controlled remotely from an iPhone or another device.³

The benefit is obvious: a research agent becomes much more useful when it can move from web information to your own files and applications. The risk is equally obvious: local file write access and native-app control create a larger blast radius for mistaken or malicious actions. The correct comparison is therefore not simply “Personal Computer has more features than Search.” It is “Personal Computer has more authority than Search.”

Macworld and Axios both framed the launch around that same trade-off, highlighting local/app context together with approval controls, auditability and kill-switch concepts.²¹²²

Perplexity's August Portable Computer launch extends the local-first idea beyond the Mac surface. Computerworld and VentureBeat report orchestration, planning, scheduling and ordinary model work running locally on Nvidia hardware, with cloud escalation reserved for steps that require current web information or stronger frontier models.²³¹⁹

The important economic point is that local work does not consume cloud token credits; the important privacy point is that local-first is not the same as local-only. Computerworld explicitly notes that cloud escalation still requires governance.²³

For most ordinary consumers this is currently a specialist product direction rather than the reason to buy Perplexity. For privacy-sensitive technical organizations, however, it could become strategically important because it separates local execution from explicit cloud escalation.

Comet turns Perplexity into an AI-native browser. The Verge's hands-on reporting describes page analysis, summaries, voice and cross-tab execution, while also finding occasional errors and slower automation in some tasks.²⁴ TechCrunch's early test similarly found useful behavior on simpler browser tasks but degradation on more complex requests, with the obvious privacy/security trade-off of broad browser access.¹⁸

For enterprise deployments, Perplexity now exposes browser policies, agent permission controls, website restrictions, action approvals and centralized management. That is a meaningful sign of product maturity because agentic browsing needs more than a consumer permission dialog.²⁵

The legal boundary is not theoretical. Bloomberg and Reuters both cover the Amazon dispute, in which courts examined whether Perplexity's shopping agent could access protected Amazon areas and who bears responsibility for automated actions.¹⁵¹⁴ This does not make Comet broadly unlawful. It demonstrates that agentic browsing can collide with third-party platform rules in ways a normal search engine does not.

Computer has its own credit economy. As of August 6, Perplexity defines 100 credits as $1. Its current examples place light tasks around 100–350 credits, complex tasks around 350–950, heavy tasks around 875–2,275 and “mega” projects around 2,400–9,800 credits.²⁰

This matters because the included credits differ sharply by plan. Consumer Pro has no recurring monthly Computer allocation; it starts with a one-time 4,000-credit promotional bonus that expires after 30 days. Consumer Max includes 10,000 credits per month plus a one-time 35,000-credit bonus. Enterprise Pro starts at 500 monthly credits; Enterprise Max at 15,000.²⁰

Ordinary Perplexity Ask searches do not consume these credits. Computer tasks do. If a user runs out, active tasks pause and resume when credits become available again. Purchased credits can be added manually or through opt-in auto-refill, and Max has a configurable spending cap.²⁰

That makes Perplexity Max's $200 headline price only part of the value equation. A heavy Computer user must think in task economics, not merely subscription price. This is one of the clearest drawbacks of the current product: a workflow can be operationally impressive while still becoming expensive or pausing if the user does not understand credit consumption.

Perplexity's consumer ladder remains straightforward at the subscription level. Standard is free. Pro is $20/month or $200/year. Education Pro is $10/month for verified students and educators. Max is $200/month or $2,000/year.²⁷²⁸

Enterprise Pro is $40 per user/month or $400/year; Enterprise Max is $325/month or $3,250/year per seat, with larger-organization discounts potentially available.²⁹

Pro is the natural reference plan for users who want advanced models, frequent research, file analysis and image generation. Max is for power users who want the highest model access and heavier use of Research/Create/Computer.²⁷

But Computer complicates the comparison because its credit system sits on top of those subscriptions. Pro does not currently include a recurring monthly Computer allocation; Max includes 10,000 credits per month. Heavy agent use can therefore turn a predictable subscription into usage-sensitive spend.²⁰

The Italian App Store also shows localized in-app purchase prices around €22/month or €229/year for Pro and roughly €229/month for Max; web pricing should remain the canonical cross-market comparison because app-store pricing can reflect local taxes and platform billing.⁵

Perplexity's current Search model table is unusually broad. Pro includes Sonar 2, GPT-5.6 Terra, Gemini 3.1 Pro, Claude Sonnet 5, Kimi K3, GLM 5.2, Grok 4.5 and Nemotron 3 Ultra. Max adds GPT-5.6 Sol and Claude Opus 5. Enterprise availability is similar but excludes Moonshot and Z.ai models and may be narrowed by administrator or retention settings.¹

This creates a methodological issue. If a user selects Claude Opus 5, the resulting output is partly evidence about Anthropic's model within Perplexity's search/retrieval layer. If the user selects GPT-5.6 Sol, OpenAI's model evidence is relevant but still does not measure Perplexity's retrieval, citation or orchestration quality. If Best or Deep Research routes models automatically, the product becomes an orchestration system rather than a single model.

AiToolMap therefore keeps third-party model benchmarks attached to the underlying model in the relevant model review rather than double-counting them as independent Perplexity performance. This is also why the stale 2025 Sonar rows from Artificial Analysis and Epoch are excluded from current scoring.¹²¹¹

Google Play currently gives Perplexity 4.6/5 from roughly 2.09 million reviews and shows more than 100 million downloads.⁶ The Italian App Store is 4.7/5 from about 32,000 ratings.⁵ These are large product-experience samples, but they measure the mobile application and service rather than controlled answer quality.

TrustRadius is smaller but more professionally oriented: 8.7/10 from 57 reviews and ratings. Its current synthesis repeatedly praises research and analysis while flagging image generation, structured-data work, accuracy/accountability and navigation as weaker themes.⁸

Gartner Peer Insights' 4.6/5 looks strong but comes from only five ratings and therefore receives very low sample confidence.⁷ PeerSpot's 1.5/5 comes from exactly one review and is essentially unusable as a population statistic.⁹

Trustpilot shows a much more negative service picture, around 1.5/5 from hundreds of reviews, with recurring complaints involving support, account switching, perceived inaccuracies and limits.¹⁰ AiToolMap does not feed that score directly into the future rating equation. It is retained only as qualitative evidence of service friction.

The combined picture is not contradictory. Large app-store populations can strongly like the product while a self-selected complaint platform concentrates users with billing or support problems. The future rating equation must preserve those population differences rather than average the stars.

Perplexity says it does not sell personal data. For Free, Pro and Max users, however, AI Data Retention is enabled by default and search/query data can be used for AI model improvement unless the user opts out in settings. The opt-out applies prospectively; previously collected training data cannot simply be removed from the training pipeline.²⁶

Enterprise is different. Perplexity says Enterprise query information is never used to train its models, uploaded files are retained for seven days, and qualifying organizations can configure custom retention and forced deletion. It also says it maintains Zero Data Retention and Zero Data Training agreements with third-party model providers for Enterprise data.²⁶

This distinction is especially important because Perplexity is multi-model. A consumer using a third-party frontier model through Search and an enterprise user operating under contractual zero-retention controls are not the same privacy surface.

The practical recommendation is simple: consumer users handling sensitive material should explicitly review and, where appropriate, disable AI data usage; organizations should evaluate Enterprise controls rather than assuming consumer settings provide the same data treatment.

Choose Perplexity if research and current web information are central to your AI use. Pro at $20 is the most sensible starting point for regular users who want advanced models, deeper research and file work. Max at $200 only makes sense when top-tier model access and heavier Computer/Research workflows justify the tenfold subscription price.

Researchers, analysts, consultants, journalists, students and knowledge workers who routinely need sources are the clearest fit. Users who already know they prefer one particular frontier model and rarely need web research may get more value from that model vendor's native product.

Computer is best evaluated as a separate buying decision. If the goal is simply answers and research, Search/Research may be sufficient. If the goal is delegated execution across files, connectors, email and scheduled tasks, Computer can justify additional complexity—but only if its credit economics and permission model fit the workflow.

Organizations handling sensitive information should evaluate Enterprise rather than extrapolating consumer privacy behavior, especially when using connected internal sources or third-party model routing.