Perplexity

Perplexity is best understood as a research-first AI workspace rather than a single chatbot or model. Its current product spans cited web Search, Advanced Deep Research, a multi-model selector, Computer for multi-step agentic work, Personal Computer for local Mac context, Comet for browser-based action, and local-first execution surfaces. Model choice remains part of the product proposition, but third-party model performance should not be mistaken for native Perplexity performance.

The September 2026 Hybrid Compute release materially changes the Mac deployment model. On eligible Apple-silicon Macs, Computer can split a task between cloud models and a local model: cloud execution handles frontier reasoning, web search and planning, while protected files and selected on-device steps can remain local. An on-device Privacy Gate classifies sensitive content and can keep it local, mask it before cloud use, refuse the step, or ask the user for consent.¹²

That is a meaningful privacy improvement over an architecture in which the useful agent must send all relevant context to the cloud, but it is not equivalent to fully offline execution. Hybrid Compute is deliberately hybrid: some task stages still use cloud models. The privacy classifier is also a control, not a guarantee; Perplexity's own PII research discusses missed detections in long-context conditions, while independent launch coverage highlights false-negative risk as a practical governance concern.²³

Hybrid Compute currently requires an Apple-silicon Mac running macOS 15 or later with at least 24 GB of unified memory. Perplexity exposes three local model choices and makes the feature available to Pro, Max and Enterprise users. Those requirements make the feature relevant to well-equipped Mac users and organizations, not to the full Perplexity installed base.¹⁴

The core research surfaces remain the most broadly useful part of the product. Search combines live retrieval with citations, while Deep Research is designed for longer source collection, document analysis and synthesis. Computer extends that workflow into execution across files, connected services and other tools. The result is unusually broad task coverage, but the product now demands more care in distinguishing which surface is acting, where data is processed and what permissions have been granted.

Perplexity still has a free Standard tier, with paid Pro and Max tiers for higher model access and advanced workflows. Computer has its own credit economics rather than behaving like an unlimited extension of a subscription. That distinction matters for frequent agent use: ordinary search and research can feel like a predictable subscription product, while delegated Computer tasks can introduce usage-sensitive cost and pauses when credits run out.⁵

Hybrid Compute improves that equation for eligible tasks because tokens generated locally do not consume cloud Computer credits. This can lower the marginal cost of local processing while simultaneously reducing data movement. It does not remove the subscription requirement or the hardware cost: the feature is available to Pro, Max and Enterprise users and requires a relatively high-memory Apple-silicon Mac.¹³

For buyers, the practical split is therefore clearer than the growing feature list suggests. Standard remains suitable for lighter cited search. Pro is the sensible mainstream paid tier for regular research and model access. Max is aimed at heavier use and broader premium access. Computer and Hybrid Compute should be evaluated as workflow and governance layers on top of those plans rather than as reasons to pay for Max by default.

Perplexity's strongest evidence continues to come from its research workflow, visible citations, broad adoption and mature product footprint. Current mobile ratings remain high: the Italian App Store shows 4.7/5 from about 32,000 ratings, while Google Play shows 4.6/5 and more than 100 million downloads. The Google Play page displays slightly different review-count totals in different sections, so the exact count should not be overstated.⁶⁷

Professional review evidence is also positive but more mixed. TrustRadius currently reports 8.7/10 from 59 reviews and ratings, with recurring praise for research, analysis and time savings and recurring weaknesses around image generation and structured-data work. Gartner Peer Insights shows 4.6/5 from only five ratings; that tiny sample is useful as qualitative enterprise context but carries little statistical weight.⁸⁹

Service sentiment is materially weaker than product ratings. Trustpilot's current profile is 1.5/5 from 778 reviews, with complaints concentrated around support, billing, subscription handling and credits. That signal is useful for service-risk context but should not be treated as a direct numeric measure of answer quality or model capability.¹⁰

Hybrid Compute adds a stronger current privacy story, but its safeguards need precise interpretation. Sensitive steps can remain local and Enterprise administrators can set organization-level sensitivity policies and review audit information. At the same time, cloud reasoning remains part of the architecture, and automated PII classification can miss sensitive material. Users handling consequential information should therefore treat the Privacy Gate as one layer of control, not as a substitute for data-governance policy.¹²³

The broader consumer/enterprise privacy distinction remains important. Consumer data settings and enterprise contractual controls are not the same surface, and teams handling confidential material should evaluate the enterprise terms and retention controls rather than extrapolating from consumer defaults.¹¹

Perplexity is a strong fit for researchers, analysts, consultants, journalists, students and knowledge workers whose tasks begin with current external information and benefit from visible sources. Users who frequently compare claims, collect evidence or turn research into documents and downstream actions get more value from the product than users who simply want a single conversational model.

Hybrid Compute particularly improves the fit for privacy-sensitive Mac workflows where local files or sensitive context need to participate in an agentic task without automatically sending every step to the cloud. The hardware floor means that advantage is concentrated among newer, higher-memory Macs and enterprise-managed Apple deployments.

Users who already prefer one frontier model and rarely need retrieval may find a model vendor's native product simpler. Likewise, organizations that do not need local context or delegated computer actions can avoid much of Perplexity's newer complexity by staying with Search and Deep Research.