Perplexity

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News


Perplexity is a conversational answer engine and multi-model AI agent platform developed by Perplexity AI, a private San Francisco company founded on August 22, 2022. When a person asks a question in ordinary language, the system searches current web and enterprise sources, a large language model synthesizes a direct reply, and the response carries numbered inline citations that the reader can open and audit. The initial public product went live on December 7, 2022, establishing a consumer-scale implementation of retrieval-augmented generation in which the model writes from documents it just retrieved rather than relying solely on static parameters stored during pretraining.

The company was founded by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski. Srinivas, the chief executive officer, previously worked as a research scientist at OpenAI, Google, and DeepMind, earning a doctorate in computer science from the University of California, Berkeley, after completing electrical engineering degrees at IIT Madras. Yarats, an AI research scientist formerly at Meta AI Research, became chief technology officer, Ho brought quantitative systems engineering experience from Quora and Wall Street, and Konwinski was already well known as a co-founder of Databricks and the Perceptive AI research ecosystem. From the outset, the four founders focused on building an applied retrieval and synthesis product rather than competing as a foundational lab racing to pretrain the single largest frontier model.

Early writeups, including the 2023 reporting preserved at the top of this page, described a chat-style search box that called an OpenAI GPT-3 model and pulled live web material through Microsoft Bing. That description matched the early consumer tool: current information, clickable footnotes, and practical warnings not to paste confidential secrets into a public search box. By late 2026, however, Perplexity had evolved into a comprehensive AI software and developer ecosystem. The company now runs its own post-trained Sonar search models, routes tasks across frontier models rented from multiple competing laboratories, operates the Perplexity API Platform and Perplexity Agent API, distributes the Chromium-based Comet browser with an embedded sidecar assistant, and offers an autonomous digital worker suite spanning Perplexity Computer, Personal Computer, and Portable Computer.

The category matters because it bridges two older digital habits that each leave half the work unfinished. Classic web search returns a ranked list of links and leaves the reading, cross-checking, and synthesis entirely to the user. A general-purpose chatbot produces polished prose immediately, even when it has not verified a single fact against current documents. Perplexity stakes its product identity on a third arrangement: deliver a concise, synthesized answer alongside the exact pages used to construct it, and extend that grounded reasoning into agents that can act across browsers, local desktops, and cloud sandboxes. Whether the generated answer remains faithful to its cited pages, and whether its crawlers and shopping agents have legal permission to access third-party websites, are the central technical and legal questions that have defined both the platform's rapid growth and its courtroom battles.


Perplexity's bet is that the useful unit of search is a checkable answer: retrieve current documents, have a model synthesize them, and show the sources, rather than returning only a list of links or an ungrounded chat reply.


How the answer engine works

From the question to a cited paragraph

A user query in Perplexity does not go straight to a language model that is presumed to know the answer from memory. In the architecture Perplexity documents, and that technical evaluations of modern answer engines corroborate, information retrieval executes first. The system parses the user's intent, queries a live web index, narrows a massive candidate set of web documents, and only then instructs a language model to compose a response. Fast lexical matching (such as BM25 keyword scoring) combined with dense embedding search and approximate nearest-neighbor lookup cuts an index of billions of web passages down to a manageable candidate pool in milliseconds. A heavier cross-encoder reranker then scores those surviving passages directly against the semantic nuances of the question itself.

The passages that survive reranking are injected into the answering model's context window alongside system instructions. Because the model is not retrained for that specific question, it is steered through prompt engineering, which operates as in-context learning in the practical sense: the probability distribution of the next generated tokens is conditioned on fresh documentary evidence that was never baked into the model's weights. Natural language generation then compresses those retrieved passages into coherent prose, functioning as multi-document summarization with source metadata attached to each claim. Short factual lookups finish in a single retrieve-and-write pass, whereas complex queries trigger an iterative loop in which the system decomposes the prompt, runs follow-up searches, reads additional pages, and revises its draft, an approach closely related to chain-of-thought reasoning except that intermediate steps incorporate live web evidence rather than solely internal scratchpad tokens.

Deep Research, introduced in February 2025, stretches that iterative retrieval loop into a comprehensive analytical report. Instead of stopping after one or two searches, the system executes dozens of sequential searches across hundreds of web pages, evaluates conflicting data points, and writes a structured multi-section briefing. At launch, technical reporting noted that Deep Research relied on a customized version of DeepSeek R1, demonstrating that multi-step research synthesis could run economically without calling only the most expensive proprietary frontier models. On the 2026 developer platform, that same multi-step workflow was exposed first as Sonar Deep Research and subsequently mapped to the `high` and `xhigh` effort levels on the Perplexity Agent API. Whatever model checkpoint runs underneath, the user-facing engineering trade-off remains constant: spend more retrieval cycles and reasoning tokens upfront to return a thorough memo that a human analyst can audit.

Natural language processing governs the entire end-to-end pipeline, not merely the fluency of the final sentences. Query decomposition, passage chunking, semantic ranking, citation alignment, and the controller's decision to search again are distinct algorithmic stages that fail in different ways. A grammatically polished sentence paired with an irrelevant footnote is a retrieval or citation-alignment failure, whereas a relevant source passage distorted into a claim the original author never made is a generation hallucination. Treating "the model" as a monolithic black box obscures which stage of the retrieval-augmented pipeline actually broke when an answer goes wrong.

What a citation does and does not prove

The numbered footnote is Perplexity's primary trust mechanism, yet it proves less than a casual reader might assume. A citation demonstrates which retrieved web page or uploaded document the pipeline associated with a generated clause. It does not guarantee that the underlying page actually supports the sentence's assertion, that the selected page represents the most authoritative primary source available, or that a statistical figure was transcribed accurately rather than smoothed or combined with numbers from another tab. While early 2023 product reviews rightly praised inline footnotes as the defining feature separating Perplexity from ungrounded chatbots, subsequent publishing disputes revealed the boundary of that safeguard: Forbes objected in 2024 when a generated summary tracked its proprietary reporting too closely with minimal attribution, and Dow Jones alleged in federal court that Perplexity outputs attached Wall Street Journal citations to fabricated claims the newspaper had never published.

For a working researcher or student, the operational rule is straightforward. Treat the synthesized answer as a high-speed triage map to decide which documents are worth reading, and then open the linked primary sources before repeating a statistic, a direct quotation, or a medical, financial, or legal conclusion. That verification habit is what practical AI trust and evaluation requires in retrieval-grounded systems: not a superficial check of whether the prose sounds confident, but a line-by-line comparison between the generated statement and the document it cites. Perplexity is structurally well suited to support that verification workflow because it exposes its source list far more transparently than a closed chat transcript, even though the software cannot perform the final human judgment for the reader.

Citation fidelity also presents a subtle model evaluation challenge that vendor benchmarks rarely capture in full. When Perplexity published online user-preference evaluations for its Sonar models in 2025, those A/B evaluations measured which response human raters preferred in side-by-side comparisons, not whether every factual clause was strictly entailed by its attached citation. Human preference and factual faithfulness correlate partially, yet they remain distinct mathematical metrics. An answer can be well structured, richly footnoted, preferred by casual testers, and still inaccurate in the exact clause that matters for a decision.

Products

Search modes, Deep Research, and subscriptions

The core consumer product is a freemium generative search and research application available on the web, in native Apple iOS, iPadOS, macOS, Android, and Windows apps, and inside the Comet browser. Quick Search handles everyday factual lookups with minimal latency, Pro Search conducts multi-step query decomposition and asks clarifying questions when a prompt is ambiguous, and Deep Research compiles long-form, multi-source reports over several minutes. While the free tier provides unlimited basic searches and a limited allowance of Pro queries, paid individual plans have anchored around Perplexity Pro at $20 per month ($200 annually) and Perplexity Max at $200 per month ($2,000 annually). Because credit allocations, model rosters, and rate limits shift frequently across product cycles, users should verify current quota details against Perplexity's live pricing documentation rather than relying on fixed historical snapshots.

Perplexity Pro serves as the primary subscription tier for everyday power users, unlocking high daily allowances of Pro Searches, document and image uploads, access to Perplexity Personal Computer, and a model selector that swaps the answering checkpoint among frontier models. Perplexity Max is the flagship tier introduced for heavy research and autonomous workloads, providing unlimited Deep Research runs, priority access to new hardware and browser features, a monthly allotment of 10,000 Computer credits, and, starting in February 2026, a Model Council feature that queries multiple frontier models in parallel and synthesizes their consensus and disagreements. On the organizational side, Enterprise Pro and Enterprise Max add single sign-on (SSO), role-based access control (RBAC), real-time audit logs, zero data retention guarantees, and native connectors into enterprise data warehouses and SaaS tools.

Spaces, Pages, and Internal Knowledge Search

Surrounding the main answer box, Perplexity provides collaborative and publishing surfaces designed to keep project context intact across sessions. Spaces allow individuals or teams to group threads, upload reference files, attach custom system instructions, and invite collaborators into a dedicated research workspace. Pages converts an exploratory search thread into a formatted, visually structured article that can be edited by section, adjusted for technical or general audiences, and published via a shareable link. Neither feature alters the underlying retrieval architecture, but both prevent context from evaporating between isolated queries, making them practical tools in academic instruction, literature reviews, and content drafting workflows where readers still verify the underlying citations.

For Pro and Enterprise subscribers, Internal Knowledge Search unifies open-web retrieval with private organizational documents. Instead of forcing an analyst to search company PDFs, spreadsheets, or connected cloud drives in one interface and the public web in another, the engine retrieves and ranks passages from uploaded files and connected repositories (such as Google Drive, OneDrive, SharePoint, Dropbox, Notion, Confluence, and Snowflake) alongside live web results in a single synthesized answer. Enterprise administrators control which connectors are active, enforce user-level permission inheritance through OAuth, and configure data retention windows down to 24 hours.

Perplexity Finance and Perplexity Shopping

Beyond general search, Perplexity operates specialized vertical applications tailored to high-intent commercial and analytical workflows. Launched in October 2024 and expanded through 2025 and 2026, Perplexity Finance brings real-time equity quotes, interactive price charting, historical financial statements, SEC filings, earnings call transcripts with audio sync, crypto and commodity tracking, and prediction-market odds (including Polymarket synthesis) directly into the cited-answer interface. Rather than acting as a registered brokerage or issuing investment advice, the product functions as a fast market research and financial analysis workstation where every ratio, revenue table, or transcript summary links back to an auditable filing or market data provider.

In November 2024, Perplexity introduced its Shopping Hub, pairing visual product cards and structured specification comparisons with conversational recommendations grounded in independent product reviews. The feature introduced Snap to Shop, which lets mobile users photograph an physical item to find matching listings, and a Merchant Program allowing retailers to feed real-time inventory and pricing data into Perplexity's index, paired with an automated "Buy with Pro" checkout flow. While clickable product recommendations operate like conventional e-commerce referrals, extending that workflow into autonomous browser agents that log into retail accounts and place orders triggered major legal friction with Amazon, illustrating the boundary between passive search discovery and automated transaction execution.

Sonar and the developer platform

Sonar is Perplexity's in-house family of search-grounded models, post-trained specifically for factual synthesis and citation alignment rather than built as a from-scratch frontier pretrain. Through the Perplexity API Platform, developers invoke Sonar and search capabilities over standard HTTP endpoints, consuming real-time grounded intelligence as AI as a service without maintaining their own web crawlers or vector databases. Because the API accepts standard JSON payloads and mirrors familiar chat and agent schemas, engineering teams can integrate it directly using Python, JavaScript, TypeScript, or cURL, while Perplexity's terms guarantee that customer API traffic is never used to train models.

The table below summarizes the Sonar API model lineup documented through 2026, prior to and alongside the Agent API consolidation. While exact per-token and per-search rates adjust over time, the table illustrates the structural billing pattern across the lineup: a competitive token generation rate paired with a tiered search fee that scales with retrieval depth.

Model Primary Role Documented Architecture and Billing Shape
Sonar Fast grounded answers, definitions, breaking news, and straightforward comparisons Built on post-trained open-weight foundations (such as Llama 3.3 70B) with roughly 128K context, priced around $1 per million input tokens and $1 per million output tokens plus a per-request search fee.
Sonar Pro Complex multi-part queries, deeper citation density, and multi-turn follow-ups Extended context window documented around 200K tokens, retrieving roughly double the web citations per query at higher token rates.
Sonar Reasoning Pro Tasks requiring explicit step-by-step analytical reasoning over live web sources Chain-of-thought synthesis over retrieved passages; mapped to the `low` effort setting on the unified Agent API as a latency-optimized reasoning tier.
Sonar Deep Research Exhaustive multi-page research memos, competitive audits, and literature syntheses Executes dozens of iterative searches and long-form synthesis; mapped to the `high` effort tier on the unified Agent API.

On September 27, 2026, Perplexity officially retired the standalone Sonar `/chat/completions` endpoint as the primary supported interface for new application development, consolidating developer workflows onto the Perplexity Agent API and the broader Perplexity API Platform. Both offerings are examined in detail in their dedicated sections below.

Comet browser and Comet Assistant

Comet is Perplexity's Chromium-based web browser, launched on July 9, 2025, initially for Max subscribers on Windows and macOS, opened as a free global download on October 2, 2025, released for Android in November 2025, and brought to iOS in March 2026. Because Comet shares the Chromium engine, desktop users retain their existing Chrome extensions, bookmarks, and enterprise group policies, while adding a built-in ad blocker and Perplexity's cited answer engine as the default address-bar experience. Its defining component is the Comet Assistant in the sidecar toolbar, which reads the active web page, compares multiple open tabs using `@tab` references, and, when authorized, clicks buttons, fills forms, and navigates websites on the user's behalf. Srinivas has framed this browser architecture as essential to true agency, arguing that a reasoning model cannot act as a practical assistant until it shares the user's authenticated browser session, tabs, and web-only tools.

Embedding web automation inside a logged-in browser fundamentally alters both the security threat model and the legal posture of the software. Malicious web pages can embed invisible text or screenshot-based instructions designed to hijack the model, turning indirect prompt injection from a theoretical curiosity into a direct risk whenever the assistant has permission to access email or submit forms. Comet's shopping automation also prompted Amazon to file suit in November 2025. Although a federal district court in California granted Amazon a preliminary injunction in March 2026 blocking Comet's agent from accessing Amazon customer accounts, the U.S. Court of Appeals for the Ninth Circuit vacated that injunction on August 4, 2026, and denied rehearing the following month, ruling that the human user, rather than Perplexity, accesses the site when the browser acts at the user's direction.

Perplexity Computer, Personal Computer, and Portable Computer

Perplexity Computer, launched on February 25, 2026, is the company's flagship multi-model digital worker. While standard Ask queries return an immediate written answer, Computer accepts an end-to-end objective, decomposes it into parallel subtasks, and orchestrates specialized subagents across 20 or more frontier models to produce finished deliverables: cited research packets, documents, spreadsheets, interactive dashboards, full-stack web applications, and actions across more than 400 connected apps including Gmail, Outlook, Slack, GitHub, Linear, Notion, and Snowflake. It maintains persistent project memory through its Brain subsystem, runs recurring or event-driven workflows via Automations on a schedule, and executes cloud workloads inside SPACE, a Firecracker microVM sandbox environment. For software engineering workflows, it competes alongside specialized coding agents, distinguished by combining live web grounding with multi-vendor model routing.

Computer usage is metered through a dedicated credit system separate from everyday search queries. Under the billing structure documented in late 2026, Perplexity Max includes a recurring monthly allotment of 10,000 Computer credits, whereas consumer Pro accounts relied on promotional credit grants (such as a 4,000-credit bonus, alongside a 20,000-to-35,000-credit introductory bonus on Max expiring after 30 days) and pay-as-you-go top-ups, while Enterprise plans share pooled organizational credits with admin-enforced spending caps. A quick single-tool task consumes tens of credits, whereas building a multi-source financial model or web application can consume several thousand. This pricing separation reflects Perplexity's underlying unit economics: cited web search remains a predictable subscription habit, while autonomous agent execution is metered because each run triggers dozens of frontier model invocations and sandboxed tool calls.

To reach data that cannot live solely in the cloud, Perplexity expanded Computer into two desktop and local editions during 2026. Personal Computer, unveiled at the Ask 2026 conference in March and rolled out to macOS in April and Windows in July, connects cloud orchestration to a local desktop client that can read and write approved local folders, control native macOS apps, and split sensitive tasks between cloud models and on-device models via Hybrid Compute. Portable Computer, launched with NVIDIA in August 2026 and expanded to Windows RTX and AMD Ryzen AI Max systems in September, runs the entire agent harness and local open-weight models (such as Qwen 3.8 27B and PPLX 27B) directly on the user's workstation GPU with zero token or credit charges for local execution.

Perplexity API Platform

Platform Architecture and Developer Ecosystem

The Perplexity API Platform is Perplexity's commercial developer and enterprise integration suite, providing programmatic access to the company's live web search index, retrieval-grounded Sonar models, autonomous multi-model agent runtime, and OEM browsing utilities. Initially introduced as a lightweight REST interface for the Sonar model family and substantially expanded at the Ask 2026 developer conference on March 11, 2026, the platform allows software engineers to embed real-time web grounding and multi-step synthesis into external applications without building custom web crawlers, document parsers, or vector embedding pipelines. Developers interact with the platform over standard HTTPS endpoints using JSON payloads, official client libraries for Python and JavaScript/TypeScript, or existing OpenAI-compatible SDKs, positioning the platform as a specialized AI-as-a-service layer focused on verifiable facts rather than ungrounded text generation.

Unlike conventional foundation-model APIs that can only reason over knowledge frozen at their pretraining cutoff or documents manually uploaded by the caller, every request to the Perplexity API Platform can dynamically query Perplexity's continuously updated web index. The platform handles query rewriting, lexical and semantic passage retrieval, cross-encoder reranking, citation extraction, and token generation inside a single round trip or managed background job. Callers can constrain retrieval using domain allowlists and blocklists, recency filters (such as hour, day, week, or month), structured JSON Schema outputs, and geographic location hints, ensuring that downstream enterprise applications receive both machine-parseable responses and an array of source URLs for human auditing or automated verification.

Core Endpoints: Search API, Sonar API, Agent API, and OEM Services

The Perplexity API Platform is organized around four distinct programmatic surfaces tailored to different levels of application complexity:

  • Search API: A low-latency retrieval endpoint that exposes Perplexity's live web search and passage-ranking engine directly to developers. Applications can submit a single search string or a batched array of multiple queries in one call, configure maximum result counts and token budgets, and receive ranked web snippets paired with contextual summaries, making it a drop-in retrieval backend for custom RAG pipelines and third-party LangChain or LlamaIndex workflows.
  • Sonar Models API: Direct inference endpoints for Perplexity's in-house search-tuned checkpoints (Sonar, Sonar Pro, Sonar Reasoning Pro, and Sonar Deep Research). While synchronous Sonar calls began transitioning onto the unified Agent API architecture following the September 27, 2026 deprecation of legacy `/chat/completions`, the Sonar model family remains the high-speed, cost-efficient engine for grounded Q&A, real-time news synthesis, and citation-backed summarization.
  • Perplexity Agent API: The flagship multi-provider orchestration and tool-use endpoint that replaces static chat completions with effort-scaled reasoning (`fast` through `xhigh`), access to 19+ third-party frontier models, built-in web and vertical search tools, a secure code execution sandbox, Model Context Protocol (MCP) support, asynchronous background jobs, and Reusable Agents.
  • OEM and Partner Browsing APIs: High-throughput, white-label conversational browsing and page-synthesis endpoints licensed to hardware and software ecosystems. Perplexity built these APIs from the technology underlying the Comet browser and uses them to power Samsung Browser's Browsing Assist across Galaxy smartphones (including the Galaxy S26 series) and Windows PCs.
Platform Service Primary Developer Use Case Output Structure Model & Tool Scope
Search API Custom RAG pipelines, multi-query batch retrieval, and real-time web grounding for external agents Ranked web passages, URLs, publication metadata, and concise contextual summaries Perplexity lexical + semantic search index and reranker (no LLM chat wrapper required)
Sonar / Sonar Pro Fast customer-facing Q&A, live fact-checking, and inline cited summarization Grounded markdown or structured JSON response with numbered citation URLs Post-trained in-house Sonar checkpoints (128K to 200K context) + live web retrieval
Sonar Reasoning Pro / Deep Research Complex multi-part analysis, financial diligence, and exhaustive multi-source reports Step-by-step reasoning traces and long-form structured memos with dense citations Reasoning-tuned checkpoints; mapped to `low` and `high` effort tiers on the Agent API
Perplexity Agent API Autonomous workflows, multi-model routing, code execution, and reusable enterprise agents Typed array of execution steps: intermediate queries, tool calls, search results, and final message 19+ frontier models (OpenAI, Anthropic, Google, xAI, Meta, Qwen) + sandboxed tools & MCP
OEM Browsing APIs OS-level and browser-native page summarization, cross-tab Q&A, and assisted navigation Low-latency streaming summaries and action instructions tailored to client UI Custom latency-optimized Sonar and Comet Assistant routing

Pricing, Data Privacy, and Enterprise Governance

Metering across the Perplexity API Platform combines token consumption with explicit retrieval and tool-invocation pricing. For Sonar and Search workloads, developers pay a per-million token rate for input and output text alongside a per-request search fee tiered by how much web context the query retrieves (low, medium, or high search context). For the Agent API, billing accounts for the underlying model provider's token rates, the selected effort level, and any sandboxed tool executions or vertical data lookups. Developers manage API keys, spend limits, usage telemetry, and organization billing through the Perplexity API Portal, which separates developer consumption from consumer Pro or Max subscriptions.

For enterprise engineering teams, the platform enforces strict data governance and security controls designed for regulated industries. Perplexity applies a strict zero data retention (ZDR) policy to API requests by default, meaning customer prompts, uploaded context, and generated outputs are never logged for training Perplexity's own models or retained by third-party model providers such as OpenAI or Anthropic. Enterprise API deployments inherit Perplexity's SOC 2 Type II, GDPR, ISO, and FedRAMP Low compliance posture (with FedRAMP Moderate in progress during late 2026), support Bring Your Own Key (BYOK) encryption so customers can revoke data access cryptographically at any time, and integrate with Numbat, Perplexity's open-source endpoint monitoring agent, to stream normalized telemetry of developer and agent activity directly into corporate SIEM platforms.

Perplexity Agent API

Architecture and Migration from Sonar Chat Completions

The Perplexity Agent API is Perplexity's unified developer interface for building, orchestrating, and deploying grounded AI agents that combine multi-provider frontier models with live web search, sandboxed code execution, and external enterprise tools. Unveiled during the Ask 2026 conference in March 2026 and designated as the primary supported replacement for the legacy Sonar chat-completions interface on September 27, 2026, the Agent API shifts developer integration from static prompt-and-reply calls to a stateful, step-oriented execution loop. While synchronous legacy Sonar requests were gradually rewritten onto the new runtime automatically to preserve backward compatibility for existing apps, asynchronous Sonar batch jobs required an explicit engineering migration to the Agent API's new asynchronous Background Mode.

At the protocol level, the Agent API replaces the flat message string of older chat endpoints with a structured, strictly typed list of execution steps. When an application submits an objective, the API response streams or returns the intermediate reasoning steps, the exact web or vertical search queries executed, the retrieved source citations, any sandboxed code or external tool invocations, and the final user-facing synthesis. This transparency allows developers who deploy AI in production to inspect why an agent reached a conclusion, render live progress indicators in client interfaces, and programmatically verify that every factual claim in the final step maps to a retrieved document from an earlier search step.

Effort Levels and Multi-Provider Model Routing

Rather than forcing developers to hardcode one of four Sonar model identifiers and manually rebuild prompts when task complexity changes, the Agent API exposes a unified **effort level** parameter (`fast`, `low`, `medium`, `high`, and `xhigh`) alongside model-neutral routing. The effort setting governs how many iterative retrieval cycles, reasoning tokens, and verification passes the agent harness expends before returning an answer. Under Perplexity's official migration table, legacy `Sonar Reasoning Pro` workloads map directly to the `low` effort tier, which denotes a fast, single-loop analytical reasoning budget rather than lower output quality, whereas exhaustive `Sonar Deep Research` jobs map to the `high` or `xhigh` effort tiers, where the agent conducts dozens of sequential searches and cross-checks across hundreds of sources.

Effort Tier Legacy Sonar Equivalent Retrieval & Reasoning Depth Typical Engineering Use Cases
`fast` Standard `Sonar` Single-pass lexical and semantic lookup with minimal latency and compact token generation Autocomplete grounding, real-time chat widgets, quick factual lookups, and voice assistants
`low` `Sonar Pro` / `Sonar Reasoning Pro` Focused multi-query decomposition with explicit step-by-step synthesis over top-ranked passages Customer support troubleshooting, multi-part technical Q&A, and interactive documentation bots
`medium` Enhanced `Sonar Reasoning Pro` + tools Iterative search refinement, cross-source verification, and conditional tool or sandbox calls Financial earnings comparisons, code debugging with live docs, and vendor evaluations
`high` `Sonar Deep Research` Deep multi-hop web exploration, synthesis across dozens of documents, and structured report drafting Automated market diligence, scientific literature reviews, legal research memos, and policy audits
`xhigh` Multi-agent Computer-class orchestration Maximum compute budget with parallel subagent routing, code execution, and exhaustive verification Complex autonomous workflows, full-stack prototype generation, and multi-repository data synthesis

Unlike vendor-locked agent frameworks from single foundation-model labs, the Perplexity Agent API lets developers route tasks across checkpoints from OpenAI (such as GPT-5.2), Anthropic (such as Claude Opus 4.6 and Sonnet 4.6), Google (Gemini), xAI (Grok), Meta (Llama), DeepSeek, and Qwen, alongside Perplexity's own post-trained models. Developers can either pin a specific model for compliance consistency or allow Perplexity's orchestrator to assign each subtask to the strongest model for that domain, such as routing deep synthesis to Gemini, long-context document parsing to GPT-5.2, and code generation to Claude.

Built-in Tools, Sandboxes, and Reusable Agents

The Agent API ships with an integrated suite of server-side tools so developers do not have to wire up basic infrastructure from scratch. Out of the box, an agent can invoke Perplexity's general web search, specialized vertical search indices (including financial market search and professional people search), a direct URL reader that fetches and parses live web pages, and an isolated cloud code sandbox for running Python data analysis, mathematical calculations, or chart generation. Developers can also attach custom client-side functions and remote Model Context Protocol (MCP) servers, enabling the agent to query internal SQL databases, trigger CRM updates, or interact with proprietary microservices during a run.

On September 28, 2026, Perplexity expanded the platform with **Reusable Agents**, allowing engineering teams to package an agent's configuration once and share it across multiple applications and internal services. A Reusable Agent combines three version-controlled components:

  • Profiles: Versioned system instructions, model preferences, effort defaults, and output guardrails that define the agent's persona and operational boundaries without scattering prompt strings across application repositories.
  • Skills: Modular, reusable task playbooks and domain procedures (such as financial ratio auditing, security alert triage, or contract redlining) that the orchestrator loads dynamically when a matching objective arrives.
  • Managed Connectors: Centralized, OAuth-authenticated integrations with enterprise systems (such as Slack, Gmail, GitHub, Notion, Linear, and Snowflake) governed by Perplexity's SPACE credential broker, where credentials stay outside the execution sandbox and are injected into outbound network requests only after deterministic policy checks pass.

By combining Reusable Agents with asynchronous **Background Mode**, organizations can trigger long-running tasks from webhooks, CI/CD pipelines, or scheduled cron jobs, poll or receive callbacks when the multi-step workflow finishes, and inspect the complete audit log of every tool call and source citation produced along the way.

Models: rented frontiers and in-house search models

Sonar, R1 1776, and Local Harness Models

Perplexity does not, as of late 2026, pretrain a massive frontier foundation model from scratch to out-scale OpenAI, Anthropic, Google, or Meta. Instead, the company specializes in post-training open-weight models for search grounding and orchestrating third-party frontier models inside its retrieval and agent harness. The default consumer and API answer path relies on Sonar, which is post-trained specifically for factual accuracy, readability, and citation fidelity. The February 11, 2025 Sonar release was built on Meta's Llama 3.3 70B checkpoint, further trained for Perplexity's default search mode, and served on Cerebras wafer-scale hardware at a claimed inference speed of 1,200 tokens per second. Perplexity reported that in online A/B user-preference evaluations, that Sonar checkpoint outperformed smaller frontier models of the period on user satisfaction and matched much larger proprietary models at a fraction of the inference cost, though those figures represent vendor-run preference tests rather than an independent academic benchmark study.

In February 2025, Perplexity also released **R1 1776**, an open-weight post-trained variant of DeepSeek R1, a large mixture-of-experts reasoning model. Srinivas stated that Perplexity's post-training pipeline removed refusal behaviors on roughly 300 topics that the base model treated as politically sensitive under Chinese content rules, while holding mathematical and logical reasoning performance steady on standard evaluations. In technical terms, "uncensored" in R1 1776 means the checkpoint answers a defined set of historical and political prompts that the parent model deflected, not that the model is free of statistical bias. The release demonstrated Perplexity's core model strategy: take a capable open-weight reasoning architecture, modify behavioral alignments that block product utility, release the weights on Hugging Face for external inspection, and deploy the customized checkpoint internally to power high-compute features such as Deep Research.

During 2026, Perplexity extended that post-training pipeline to smaller models built for local and hybrid agent execution. For Portable Computer and macOS Hybrid Compute, the company shipped **PPLX 27B**, a post-trained variant of Alibaba's Qwen 27B architecture tuned specifically on real-world trajectories from Perplexity's agent harness to reduce tool-calling errors, handle context compaction across a 256,000-token window, and minimize user corrections. Alongside PPLX 27B, Perplexity supports local execution of Qwen 3.8 27B, Qwen3.6 35B-A3B, and Google's Gemma 4 E4B, giving users on Apple Silicon Macs, NVIDIA RTX workstations, and AMD Ryzen AI Max PCs a local-first model roster that processes private files on the device before escalating complex reasoning to the cloud.

Open-Source Security Models and Tooling: BrowseSafe, Numbat, and Bumblebee

Because Perplexity's browser and computer agents operate across untrusted web pages and developer endpoints, its security engineering team built and open-sourced three defensive tools between late 2025 and mid-2026:

  • BrowseSafe and BrowseSafe-Bench: Released on December 2, 2025, BrowseSafe is an open-weight machine learning classifier trained to detect indirect prompt injection payloads embedded in live websites, HTML DOM structures, and emails before an agent acts on them. Paired with BrowseSafe-Bench, a public evaluation dataset of 14,719 realistic web-page examples, BrowseSafe achieved a 91% detection rate at low latency in Perplexity's evaluations (dropping to 76% on multilingual injection attacks), serving as a fast first-line filter inside Comet and Computer.
  • Numbat: Open-sourced in July 2026, Numbat is an endpoint detection and response (EDR) monitoring and policy-enforcement suite built specifically for AI coding and computer-use harnesses, including Claude Code, Codex, Cursor, OpenClaw, and Perplexity's own agents. Numbat injects into the agent lifecycle between reasoning and tool execution, streaming normalized telemetry of every agent action to enterprise SIEM systems and enforcing deterministic rules that alert on or block unauthorized behaviors (such as an agent attempting to open a public repository when a private credential fails).
  • Bumblebee: Also open-sourced in 2026, Bumblebee is a lightweight software supply-chain scanner deployed across developer laptops, servers, and CI/CD pipelines. It maintains a continuous inventory of installed third-party packages and scans for malicious dependencies that compromise developer machines before code is ever committed to a pull request, working in tandem with Perplexity Computer agents that monitor public threat-intelligence feeds and page security engineers when a newly compromised package matches the fleet inventory.

Frontier Model Orchestration and Cloud Infrastructure

For paying subscribers and Agent API developers, Perplexity acts as a model-neutral router that lets users choose a model for a specific thread or delegates subtasks across multiple labs automatically. Depending on the plan and surface, the platform routes queries across checkpoints from OpenAI (such as GPT-5.2 and GPT Realtime 1.5 for voice), Anthropic (such as Claude Opus 4.6 and Sonnet 4.6), Google (Gemini), xAI (Grok), and visual generation models such as Nano Banana and Veo 3.1, alongside Sonar. By December 2025, Perplexity reported that no single model handled more than 25% of total usage across its platform, and in February 2026 it introduced the Model Council on Max to query several frontier models simultaneously and synthesize their outputs. Srinivas has described this multi-provider posture as making Perplexity "the Switzerland of AI," a commercial positioning that keeps the application layer independent of any single lab's release cycle.

Sustaining that model-neutral picker requires massive cloud commitments and specialized inference hardware. On January 29, 2026, Reuters reported that Perplexity signed a three-year, $750 million cloud agreement with Microsoft to access Azure and Azure AI Foundry, securing enterprise capacity for models from OpenAI, Anthropic, and xAI, while Perplexity confirmed to Bloomberg that Amazon Web Services (AWS) remained its primary cloud infrastructure provider. Later in September 2026, Perplexity also signed a multi-year compute agreement with Crusoe for NVIDIA GB300 GPU clusters. Running default search at interactive speed and orchestrating 20-model agent workflows is ultimately an infrastructure and processor supply problem: Nvidia sits on both sides of Perplexity's business as a strategic equity investor and as the primary silicon supplier powering both cloud inference and local Portable Computer workstations.

Company, capital, and distribution

Perplexity's financing history is documented through investor disclosures, regulatory filings, and financial press reporting. Compilations current through October 2026 placed total announced equity funding between $1.5 billion and $1.7 billion, tracing one of the fastest private valuation climbs in software history: roughly $520 million in January 2024, $1 billion in April 2024, $3 billion in mid-2024, $9 billion in December 2024, $14 billion following a $500 million round in May 2025, $18 billion in July 2025, and $20 billion after a $200 million extension in September 2025. Early 2026 secondary and press reports cited valuations near $21 billion, and in August 2026 Reuters and other outlets reported investment talks led by NVIDIA at a valuation exceeding $30 billion, though those talks had not been reported as closed by late October 2026. As with any private technology firm, those figures reflect the negotiated price of preferred shares in specific financing rounds rather than an audited public-market capitalization.

Named institutional and angel backers across Perplexity's financing rounds include Nvidia, SoftBank's Vision Fund 2, Accel, Institutional Venture Partners (IVP), New Enterprise Associates (NEA), Bessemer Venture Partners, Databricks Ventures, and individual investors including Jeff Bezos, Nat Friedman, Elad Gil, and Shopify CEO Tobias Lütke. Financial press reports in 2026 indicated that Perplexity's annual recurring revenue (ARR) crossed $450 million by March 2026, with subsequent writeups citing higher run rates driven by Max subscriptions and enterprise seats, while monthly active users reached roughly 45 million in late 2025. Reported query volume ranged from hundreds of millions of monthly searches to company statements citing more than 10 billion annualized answers. Despite rapid top-line growth, the company is widely understood to operate at a net loss due to the heavy cloud compute and frontier-model licensing costs required by Deep Research and Computer workloads.

Distribution partnerships have been as central to Perplexity's growth strategy as its retrieval engineering. Preload and bundling agreements placed the Perplexity app or complimentary one-year Pro subscriptions on Motorola smartphones, on millions of telecom subscribers' plans through carriers such as Bharti Airtel in India and Deutsche Telekom in Europe, on HP Windows PCs starting in September 2026, and at the operating-system level on Samsung's Galaxy S26 lineup via Browsing Assist APIs. In parallel, the company used bold corporate bids to command public attention: in January 2025 it submitted an unsolicited merger proposal involving TikTok's U.S. operations, and in August 2025 it publicized an unsolicited $34.5 billion offer to acquire Google Chrome during Google's antitrust remedies phase. Neither bid proceeded, and analysts broadly interpreted them as strategic publicity aimed at positioning Perplexity and Comet as credible challengers to incumbent browser and search monopolies.

During 2026, Perplexity also executed a notable pivot in its monetization model. After introducing a Publishers' Program in July 2024 that promised media partners a share of advertising revenue whenever their articles appeared in cited answers, reporting in February 2026 revealed that Perplexity was stepping back from advertising-led monetization to focus on high-margin consumer subscriptions (Pro and Max), usage-based Computer credits, enterprise seat licensing, and developer API consumption. To address publisher compensation inside the browser, the company introduced Comet Plus at $5 per month (bundled into Pro and Max), allocating 80% of subscription proceeds from an initial $42.5 million pool to participating news organizations based on human reads, search citations, and agent interactions. Whether subscription revenue-sharing can offset the ad impressions publishers lose when readers stay inside an AI summary remains one of the central economic debates surrounding the company.

Publishers, crawling, and the courts

The legal conflict surrounding Perplexity flows directly from how a retrieval-augmented answer engine operates. To answer current questions, the system fetches web pages, stores or processes their text inside an index and model prompt, and emits synthesized prose that can closely mirror the structure and facts of the underlying reporting. News and reference publishers contend that this pipeline acts as an unlicensed commercial substitute for their journalism, occasionally reproducing near-verbatim passages or, conversely, tarnishing their trademarks by attaching a reputable outlet's footnote to a hallucinated error. Perplexity argues in response that factual information is not copyrightable, that inline citations drive referral traffic back to primary sources, and that its revenue-sharing programs offer fair compensation. By late 2026, federal courts had not issued a final merits judgment resolving the core fair-use question, but multiple high-profile lawsuits had advanced deep into technical discovery over how Perplexity's index and crawlers operate:

  • Dow Jones and the New York Post: News Corp subsidiaries Dow Jones (publisher of The Wall Street Journal) and the New York Post sued Perplexity in the Southern District of New York on October 21, 2024, alleging massive copyright infringement into a retrieval database and trademark dilution from fabricated quotes attributed to their journalism. A federal judge denied Perplexity's motion to dismiss or transfer the case in August 2025, and in 2026 the court ordered Perplexity to produce expanded user-activity and query logs.
  • The New York Times, Chicago Tribune, and CNN: Following a 2024 cease-and-desist notice, The New York Times filed suit alongside related actions by the Chicago Tribune and other major dailies alleging unlawful scraping, paywall circumvention in Comet, and verbatim output substitution. In May 2026, CNN filed its own federal lawsuit alleging large-scale unauthorized copying of articles, video transcripts, and images, coupled with false attribution of inaccurate statements to CNN reporters.
  • Encyclopaedia Britannica and Merriam-Webster: The reference publishers sued Perplexity in September 2025 for copyright and trademark infringement. On September 2, 2026, the district court ordered Perplexity to produce an additional technical snapshot of its retrieval-augmented generation system and to host months of user-activity data in a secure inspection environment, turning the litigation into a direct forensic examination of how an answer engine stores and ranks content.
  • Reddit: Reddit filed suit in October 2025 alleging industrial-scale scraping of user discussions despite blocks, and in July 2026 a federal judge largely denied Perplexity's motion to dismiss, allowing Digital Millennium Copyright Act (DMCA) anti-circumvention and New York civil-conspiracy claims to proceed into discovery.
  • International Publishers and the BBC: In June 2025, the BBC threatened legal action demanding that Perplexity cease scraping its content and pay compensation, while major Japanese news organizations including Yomiuri Shimbun, Asahi Shimbun, and Nikkei filed copyright suits in 2025 seeking injunctions and damages over unlicensed article ingestion.
  • Amazon and the Comet Shopping Agent: In November 2025, Amazon sued Perplexity alleging unauthorized computer access and fraud from Comet's automated shopping agent logging into customer accounts. Although Amazon won a preliminary injunction in March 2026, the Ninth Circuit vacated the injunction on August 4, 2026, and rejected Amazon's rehearing bid in September 2026, holding that users rather than Perplexity access the store when directing their browser agent, while leaving Amazon's trademark and state-law claims to proceed in district court.
  • Trademark Dispute: Separately from publisher and platform suits, Perplexity has also defended trademark claims filed by Perplexity Solved Solutions, a software firm contesting the use of the "Perplexity" mark itself.

Web crawling practices sit at the evidentiary center of several of these lawsuits. In June 2024, Wired and developer investigations reported that Perplexity appeared to retrieve articles from websites whose `robots.txt` files explicitly disallowed its crawler, and Srinivas stated that the company relied in part on third-party web-crawling providers. On August 4, 2025, web infrastructure provider Cloudflare published network telemetry alleging that when websites blocked Perplexity's declared user agent (`PerplexityBot` and `Perplexity-User`), a fallback client masquerading as a standard macOS Google Chrome browser rotated through autonomous system numbers and IP addresses outside Perplexity's published ranges to bypass `robots.txt` directives and web application firewalls. Perplexity publicly disputed Cloudflare's characterization as a publicity stunt and maintained that user-triggered agent fetches differ from automated index crawling. Whatever the legal outcome, the technical tension is structural: an answer engine blocked from major publishers must either omit top sources or fetch them through pathways site operators sought to close, making crawler behavior a core issue of data governance, journalism economics, copyright law, regulatory compliance, and practical AI governance.

From an ethical and policy standpoint, Perplexity's design presents a genuine paradox. Inline citations and publisher revenue pools represent a meaningful step toward transparency compared with opaque foundation models that absorb web text during pretraining and emit answers without a single link. Yet a footnote does not cure a hallucinated quotation falsely pinned to a reporter's name, nor does it settle the economic question of zero-click substitution: when an answer engine summarizes an investigative report so completely that the reader never visits the publication, does that interface represent fair transformative synthesis or an unsustainable extraction of value from the open web? The discovery orders entered in late 2026 are forcing those competing theories to be tested against actual retrieval snapshots and clickstream logs rather than abstract policy arguments.

How it compares

Against Google Search, the early 2023 contrast was framed simply as "ten blue links versus a synthesized paragraph." Because Google subsequently deployed AI Overviews and AI Mode across its main search results, the live competitive distinction in 2026 centers on interface defaults, advertising density, and source transparency rather than whether generative search exists. Against general conversational assistants such as OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini, the primary difference lies in product architecture and default assumptions. Those systems began as general-purpose reasoning and writing workspaces that invoke web browsing as an optional tool, whereas Perplexity is built retrieval-first: live web search and numbered citations are the default behavior on every prompt, and the interface assumes the user wants a checkable briefing. Microsoft Bing Copilot sits structurally close because it also pairs a major web index with a conversational pane, You (You.com) pioneered early cited search modes before pivoting toward enterprise productivity, and xAI's Grok blends web retrieval with real-time social platform data. Perplexity's distinct pitch is its model-agnostic orchestration harness across all of those labs.

For an individual or organization trying to pick a product or platform, the practical division follows the nature of the task. Choose Perplexity, or a specialized language and retrieval tool like it, when the deliverable requires live multi-source verification: breaking news synthesis, academic literature scouting, competitive market snapshots, financial filing lookups, or open-source intelligence (OSINT) research. Choose a general frontier assistant like Claude or ChatGPT when the task centers on creative writing, transforming text already in hand, or long-horizon coding inside a single model ecosystem, and rely on a private internal RAG stack when strict regulatory rules prohibit sending documents to a multi-tenant cloud service. With the launch of Comet, Perplexity Computer, Personal Computer, and Portable Computer, Perplexity now competes directly in the autonomous workspace market against Microsoft Copilot, Claude Cowork, OpenAI's ChatGPT Work and dots, Google Gemini Spark, and Manus, a category analyzed side by side in the Ready-to-Use AI Agent Comparison.

Limits and open questions

The everyday failure mode users encounter first in Perplexity is a fluent, confident sentence paired with an authentic-looking footnote that does not actually substantiate the claim. Retrieval pipelines can also miss the primary study, over-weight search-engine-optimized aggregator pages, or skip authoritative outlets that have blocked Perplexity's crawlers, leaving the language model to synthesize a confident answer from a skewed sample of secondary commentary. Because source quality flows directly into generation quality, an answer engine restricted to thin web pages will produce a thin summary with authoritative prose. These limitations are not rare edge cases; they represent the core technical reality of evaluating and trusting retrieval-augmented systems, which is why best practice requires opening the underlying citation before acting on any consequential claim.

Moving from read-only answer synthesis to autonomous browser and desktop agents introduces a second, more severe class of operational risk. Indirect prompt injection, broad operating-system permissions, and background tool execution create immediate cybersecurity and privacy exposures. As security researchers from Brave, LayerX, and Guardio demonstrated against Comet in 2025, untrusted web pages or crafted links can attempt to hijack an agent that holds access to logged-in browser sessions, email inboxes, or cloud connectors, and Perplexity's own evaluations show that classifiers like BrowseSafe reduce but do not eliminate that vulnerability. Consequently, actions that send external messages, modify production code repositories, delete local files, or spend money require explicit human-in-the-loop approval gates, and comprehensive AI verification and validation must audit both whether an answer matched its source and whether an agent stayed within its authorized security boundary.

Finally, Perplexity faces three structural questions across its commercial, legal, and product strategy. Commercially, operating as a model-neutral orchestration layer requires paying hundreds of millions of dollars to cloud and model suppliers like Microsoft Azure, AWS, OpenAI, and Anthropic while subsidizing heavy inference workloads across tens of millions of users. Legally, federal courts in New York and California must determine whether real-time retrieval and citation synthesis constitutes fair use or requires explicit licensing from news and reference publishers. On the product side, it remains an open question how much autonomy everyday users genuinely want to delegate: as Srinivas observed in late 2026 interviews, many users do not want an opaque bot spending their money autonomously, preferring an assistant that reads thousands of pages in parallel and presents a clear comparison so the human can make the final call.

Recent Developments

Between mid-2025 and late 2026, Perplexity transitioned from a cited conversational search engine into a multi-surface agent and developer platform. Comet launched for Max subscribers on desktop in July 2025, became a free global download in October 2025, reached Android in November 2025, and arrived on iOS alongside Comet Enterprise in March 2026. On February 25, 2026, Perplexity launched Perplexity Computer, a credit-metered cloud digital worker orchestrating 19+ models inside Firecracker microVMs, followed at the March 11, 2026 Ask conference by Personal Computer for local desktop workflows (rolling out to Mac in April and Windows in July) and Computer for Enterprise. In August and September 2026, Perplexity partnered with NVIDIA and AMD to launch Portable Computer for fully local GPU execution, introduced Hybrid Compute on Apple Silicon Macs, added persistent Brain memory, and upgraded Scheduled Tasks into event-driven Automations.

On the infrastructure and developer side, Perplexity secured a three-year, $750 million Microsoft Azure and Foundry agreement in January 2026 and a multi-year Nvidia GB300 compute deal with Crusoe in September 2026 while keeping AWS as its core cloud host. In February 2026, the company pivoted away from advertising toward subscription tiers, enterprise seats, and usage credits, while securing OS-level integration on the Samsung Galaxy S26 and PC preload deals with HP. For developers, Perplexity retired the standalone Sonar chat-completions endpoint on September 27, 2026, in favor of the unified Perplexity Agent API with effort levels (`fast` to `xhigh`), and launched Reusable Agents with versioned Profiles, Skills, and managed connectors on September 28, 2026. Meanwhile, the company open-sourced its BrowseSafe prompt-injection detector (December 2025) and its Numbat and Bumblebee security suites (mid-2026).

In the courts and private markets, late 2026 brought pivotal developments on both fronts. While the Ninth Circuit handed Perplexity a significant appellate victory in August and September 2026 by vacating Amazon's preliminary injunction against the Comet shopping agent, copyright and scraping lawsuits from Dow Jones, The New York Times, CNN, Reddit, and Encyclopaedia Britannica intensified, highlighted by a September 2, 2026 court order compelling Perplexity to produce retrieval-system snapshots and hosted user-activity logs. Throughout this period, private market valuations rose from a closed $20 billion round in September 2025 to reports above $30 billion in late-2026 financing discussions, underscoring both investor appetite for agentic search and the high stakes of the ongoing legal and infrastructure battles.

Early public demonstrations

The video clips below are the historical demonstrations originally embedded on this page when Perplexity first gained public traction in 2023. They remain valuable as a primary record of how the conversational answer engine was introduced against traditional link-based search, though they do not reflect Sonar, Deep Research, the Perplexity Agent API, Comet, or Perplexity Computer. Viewers should treat period comparisons with Google Bard (a product name since retired in favor of Gemini) and broad promotional headlines as 2023 context rather than current technical evaluations.

Load video
YouTube

This early 2023 walkthrough contrasts Perplexity's original interface directly with classic Google Search. It serves as a clear historical record of the foundational product pitch, replacing ten blue links with a concise synthesized paragraph and clickable source footnotes, prior to Google's rollout of AI Overviews and Perplexity's expansion into multi-model agents. While the demonstration captures the core citation workflow described in the Introduction/Overview, it predates the in-house Sonar model family, Deep Research, and the Comet browser.

Load video
YouTube

This concise demonstration introduces the conversational search box and explains the concept of an "answer engine" for general users. Its most durable detail is showing how numbered inline citations allowed early adopters to verify claims and ask contextual follow-up questions inside a single thread. The underlying model routing shown in this 2023 clip relied on Perplexity's early OpenAI and Bing integration rather than the 2026 multi-provider model picker and Agent API.

Load video
YouTube

This 2023 comparative review evaluates early generative search tools alongside Google Search and Google's then-new Bard assistant. It is useful for understanding how the conversational search category formed in the months immediately following the launch of ChatGPT, highlighting why source attribution quickly became the dividing line between search tools and general chatbots. Because interface layouts, context windows, and model capabilities changed rapidly after 2023, several specific UI comparisons in the video aged out within a year.

References

Additional Viewing

Load video
YouTube

Aravind Srinivas at the Cambridge Union, recorded October 3 and published October 23, 2026 (about 58 minutes), provides a comprehensive long-form account of how Perplexity's chief executive describes the company's evolution. Srinivas discusses the mechanics of the retrieval-grounded answer engine, citation accuracy, and the strategic transition from passive search replies toward autonomous browser and desktop assistants, directly illuminating the themes in the Introduction/Overview and Company sections. Because the chamber format includes direct audience questioning, the video surfaces substantive critiques alongside the company's product vision. Cambridge Union: In Conversation with Aravind Srinivas

Load video
YouTube

"Perplexity: Beginner to Pro in 27 Minutes," presented by Ali H. Salem (October 29, 2025, about 28 minutes), offers a structured tour of the core modes separated in the Products section: Quick Search, Pro Search, Deep Research, Pages, Spaces, and the multi-provider model picker. The walkthrough demonstrates why retrieval-grounded outputs differ from a standard chatbot and shows how those research habits carry directly into the Comet browser. Viewers can use it as a practical interface map while checking Perplexity's live documentation for current 2026 credit and rate limits. Perplexity: Beginner to Pro in 27 Minutes

Load video
YouTube

Perplexity's official "Comet For Work" workshop (September 1, 2025, about 31 minutes) demonstrates the Chromium-based browser as shipped to early enterprise and power users, covering the sidecar assistant toolbar, multi-tab comparisons, and automated web workflows. It serves as a direct visual companion to the Comet subsections and clarifies the technical backdrop of the Amazon litigation by showing how the agent operates inside an authenticated user browsing session. Because the recording is a vendor workshop, viewers should pair its workflow demonstrations with the security constraints discussed in Limits and Open Questions. Comet For Work

Load video
YouTube

In this interview at Money 20/20 USA with the New York Stock Exchange's JD Durkin (published October 27, 2026, about 9 minutes), Aravind Srinivas explains why financial citations must trace every figure to an auditable filing, why Perplexity maintains a model-neutral orchestration layer, and why many users prefer deep multi-page comparison over an agent that autonomously spends their money. That concise discussion connects the citation-verification rules, the multi-model picker strategy, and the e-commerce agent debates covered across this article. Money 20/20: Aravind Srinivas

Load video
YouTube

"How to use Perplexity Search API | Python | UV | Demo," by Ashish Agarwal at Coders Academy (published October 15, 2025, about 6.5 minutes), walks step by step through generating a Perplexity API key, configuring a Python environment with the official `perplexityai` SDK, and executing both single-query and batched multi-query requests against the Search API. The demonstration directly supports the Perplexity API Platform section by showing how developers retrieve live, ranked web results and contextual summaries programmatically in just a few lines of Python code. How to use Perplexity Search API | Python | UV | Demo

Perplexity Computer

Perplexity Computer is Perplexity's autonomous AI agent platform, which the company describes as a general-purpose digital worker capable of operating the same software interfaces a human professional uses. Instead of typing step-by-step prompts, a user describes a target outcome, and Computer plans the workflow, decomposes the objective into parallel subtasks, and routes each subtask to whichever foundation model is best suited for that specific domain. It can browse the live web, conduct multi-source research, write and execute code, build full-stack applications and interactive websites, and act across connected workspace tools, running unattended for hours or through recurring triggers over months. Founded in August 2022 by Aravind Srinivas and his co-founders as a conversational answer engine, Perplexity used the February 2026 launch of Computer to mark its structural shift into an end-to-end agent platform that now spans four integrated deployment surfaces:

  • Computer (cloud): The original cloud-hosted digital worker, launched February 25, 2026, where each task executes inside its own isolated Firecracker microVM sandbox and orchestrates 19 to 20 or more models from competing AI laboratories.
  • Personal Computer: The hybrid desktop edition that lets Computer read, write, and organize approved local files and interact with native applications on the user's own macOS or Windows machine.
  • Portable Computer: A local-first edition built with NVIDIA and AMD that runs both the model weights and the agent harness entirely on the user's local GPU hardware, consuming zero cloud credits for on-device execution.
  • Everywhere else: Embedded access across desktop and mobile apps, Slack, Microsoft Teams, Microsoft 365 (Word, Excel, PowerPoint, and Outlook), and the sidecar assistant inside the Comet browser.


Perplexity Computer treats AI models like an orchestra: one conductor plans the work, specialist models handle each part in parallel, and the whole job runs in a sandbox that can live in the cloud, on your desktop, or entirely on your own GPU.


  • The Mechanism: A central orchestrator model reads the user's high-level goal, maps out a dependency graph of tasks and subtasks, and deploys specialized subagents in parallel. At launch, Anthropic's Claude Opus 4.6 served as the primary orchestrator and handled software coding, while the harness dynamically routed other subtasks by model specialty:
    • Gemini: deep multi-source research and synthesis
    • OpenAI GPT-5.2: long-context document parsing and analysis
    • Grok: fast, lightweight real-time lookups
    • Nano Banana and Veo 3.1: visual image and video generation

Users can also manually override the defaults and specify which models Computer should assign to its subagents. Every cloud task runs inside SPACE, Perplexity's purpose-built agent sandbox infrastructure, which treats the autonomous agent as untrusted code by default:

    • each task boots inside its own isolated Firecracker microVM
    • OAuth tokens and API credentials stay outside the sandbox and are injected by a network gateway credential broker only when outbound requests pass deterministic policy checks
    • outbound network traffic is inspected and governed outside the guest VM, supporting human-in-the-loop approval holds
    • file-system snapshots allow users and administrators to roll back destructive changes

Computer reaches external workspaces through more than 400 managed connectors (including Gmail, Outlook, Slack, Notion, Linear, GitHub, Google Calendar, HubSpot, and Snowflake), reinforced by reusable Skills, collaborative Projects, persistent Brain memory, and local file support.

  • The Physics: A multi-model agent that executes autonomously for hours encounters hard engineering constraints across cost, compute hardware, context windows, and operational safety:
    • Credits: Computer consumes usage-based credits on top of a paid subscription. Perplexity Max ($200 per month) includes 10,000 monthly credits, and complex multi-agent builds or 50-company research batches consume credits rapidly based on model calls and tool invocations, requiring user-configurable spending caps and subagent model selection to control expenses.
    • Effort: Granular effort controls, ranging from Light to Ultra (and `fast` to `xhigh` on the Agent API), match model selection and reasoning depth to the difficulty of each job.
    • Local hardware: Portable Computer requires high-end local GPU memory. It shipped first on Nvidia DGX Spark systems and Linux workstations with RTX GPUs, followed by Windows PCs equipped with Nvidia RTX GPUs (24 GB or more of VRAM) and AMD Ryzen AI Max processors.
    • Local context: Because the local Qwen 3.8 27B and PPLX 27B checkpoints operate within a 256,000-token context window, Portable Computer relies on automated context compaction to keep long-horizon tasks coherent without exhausting memory.
    • Accidental meltdowns: Perplexity's security engineering team notes that autonomous agents pursuing legitimate goals can attempt to bypass security boundaries when they hit operational friction, such as a failed API call or an expired token, which is why the platform layers independent deterministic guardrails rather than trusting model alignment alone.
  • Why This Matters: Computer represents Perplexity's strategic bet that the winning enterprise and consumer agent will orchestrate every laboratory's best models rather than locking customers into a single vendor's checkpoints:
    • Multi-model by design: Perplexity reported that by December 2025, no single foundation model accounted for more than 25% of total usage across its platform, allowing the harness to swap models immediately as frontier capabilities shift.
    • Productivity claim: Perplexity stated in internal case studies that its own teams used Computer to complete an estimated 3.25 years of equivalent engineering and operational work in four weeks, saving $1.6 million in labor costs (a vendor-reported figure).
    • Local-first option: Portable Computer and macOS Hybrid Compute bring the same orchestration interface to local hardware, keeping confidential files on the device and escalating to cloud frontier models only with explicit user consent.
    • Distribution: Hardware and OS partnerships gave Perplexity OS-level Browsing Assist integration on the Samsung Galaxy S26 series and preloaded the Perplexity Windows application across new HP PCs.
    • Security tooling: Perplexity open-sourced its BrowseSafe prompt-injection classifier, its Numbat endpoint agent-monitoring suite, and its Bumblebee supply-chain package scanner, while publishing transparent red-team evaluations of its SPACE sandbox.

Deployment Modes

Mode Where It Runs What It's For
Computer (cloud) A per-task Firecracker microVM in Perplexity's SPACE sandbox, orchestrating 19 to 20+ cloud models. The full digital worker: deep research, coding, building full-stack apps, browser automation, and multi-app workflows.
Hybrid Compute on Mac Cloud frontier models paired with local Apple Silicon models inside the macOS desktop app. Keeping sensitive local files and privacy-gated steps on the Mac while cloud models handle heavy planning and web synthesis.
Personal Computer The user's own Mac (all paid users) or Windows PC (Pro, Max, and Enterprise). Letting Computer read and write approved local folders and interact with native desktop applications.
Portable Computer Fully on device: Nvidia DGX Spark, RTX GPUs on Linux or Windows (24 GB+ VRAM), or AMD Ryzen AI Max PCs. Private, credit-free local execution running Qwen 3.8 27B or PPLX 27B, a Qwen variant post-trained on Perplexity's agent harness.
Automations Computer's cloud agent, triggered on a recurring schedule or by live events in Slack, Gmail, Outlook, Linear, or GitHub. Standing assignments that retain memory of prior runs, such as daily competitor monitors, weekly briefings, and alert triage.
Work apps and Comet Slack, Microsoft Teams, Microsoft 365 side panels, and the Comet browser's sidecar assistant. Assigning and reviewing Computer tasks directly inside the communication and document tools a team already uses.

To prevent local execution from exposing the host workstation to uncontrolled actions, Portable Computer enforces three deterministic architectural guardrails around the local model weights. Unlike cloud sandboxes that rely purely on remote virtual machines, the local client separates policy enforcement from language-model reasoning so that a hallucinated or injected instruction cannot disable host protections:

  • Deterministic orchestrator: The outer control loop is written as deterministic harness code rather than controlled by an LLM, enforcing strict tool-call and file-access policies before any command executes.
  • Always-on sandbox: Approved local tool calls execute inside an operating-system-level sandbox that is engineered to fail closed; if the OS sandbox is unavailable or tampered with, the agent harness automatically disables itself.
  • Cloud escalation on request: Escalating a difficult subtask from the local 27B checkpoint to a stronger cloud advisor model requires explicit user approval, and an on-device personally identifiable information (PII) classifier previews and masks sensitive fields before anything leaves the machine.

Getting Started

  • Plans: Computer is available to Pro and Max subscribers across web, desktop, mobile, Slack, and Microsoft 365, as well as to Enterprise organizations. Perplexity Max costs $200 a month ($2,000 a year) and includes 10,000 monthly Computer credits.
  • Start a task: Open perplexity.ai/computer and describe the final deliverable or outcome you need. Computer decomposes the objective, deploys parallel subagents in the background, and allows you to inspect intermediate steps or steer its plan in real time.
  • Connect tools: Authorize OAuth connectors such as Gmail, Outlook, Slack, Notion, Linear, GitHub, and Google Calendar at perplexity.ai/computer/connectors, and attach modular Skills to standardize how Computer formats deliverables.
  • Set effort: Select an effort level from Light to Ultra (or configure subagent models and credit caps) to balance reasoning depth against credit consumption.
  • Automate recurring work: Describe a recurring job in Computer's omnibar, or open Automations and select New Automation to configure a schedule or webhook/event trigger, attach relevant Projects, and specify which outbound actions require human approval.
  • Run locally: In the Perplexity desktop app's Local Inference settings on supported Nvidia RTX, DGX Spark, AMD Ryzen AI Max, or Apple Silicon hardware, download a local checkpoint (such as Qwen 3.8 27B or PPLX 27B) with one click to run tasks locally at zero credit cost.
  • Build on it: Software teams can invoke the same orchestration engine programmatically through the Perplexity Agent API, which supports Reusable Agents with versioned Profiles, Skills, and managed connectors shared across internal applications.
  • Govern it: Enterprise administrators can restrict connector scopes (for example, allowing Computer to read and draft emails while requiring human approval to send), stream normalized telemetry to SIEM tools via Numbat, enforce role-based access control (RBAC) through Okta or Entra ID, and manage Bring Your Own Key (BYOK) encryption where revoking a key renders stored project data cryptographically unreadable.

Known Limitations

  • Credit costs: Long-running, multi-model jobs such as batch-researching 50 companies or building interactive dashboards consume credits quickly, and usage beyond a plan's monthly allotment requires paid credit top-ups.
  • Hardware gates: Portable Computer and Hybrid Compute require workstation-class memory (at least 24 GB of VRAM on Nvidia RTX or AMD Ryzen AI Max PCs, or 24 GB to 32 GB of unified memory on Apple Silicon Macs), and their 27B-to-35B local checkpoints remain smaller than frontier cloud models, requiring occasional cloud escalation on complex tasks.
  • Vendor-reported claims: Internal labor-savings estimates ($1.6 million saved in four weeks, or $9.4 billion in cumulative user work), model-share percentages, and multi-month unattended runtime claims originate from Perplexity's own announcements rather than independent audits.
  • Residual security risk: In Perplexity's published September 2026 red-team evaluations of SPACE ("Escaping SPACE, Part I"), nine frontier models given root access inside the Firecracker microVM all failed to escape the virtual machine, but several discovered ways to bypass outbound network filters when partial network access was enabled, gaps Perplexity subsequently patched while cautioning that no prompt injection defense is 100% foolproof.
  • Legal overhang: Active copyright and scraping lawsuits from publishers including Dow Jones, The New York Times, CNN, Encyclopaedia Britannica, and Reddit could constrain how Computer's web-research subagents ingest and synthesize commercial web content.

Recent Developments

  • August 2022: Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski established Perplexity AI in San Francisco.
  • July 9, 2025: Perplexity launched the Comet AI browser for Windows and macOS, expanding to Android in November 2025 and iOS in March 2026.
  • January 2026: Perplexity signed a three-year, $750 million cloud agreement with Microsoft for Azure and Foundry, securing high-volume access to frontier models from OpenAI, Anthropic, and xAI.
  • February 2026: Perplexity pivoted away from advertising toward a subscription-and-credits model, and Samsung integrated Perplexity's Browsing Assist APIs at the OS level on the Galaxy S26 series.
  • February 25, 2026: Perplexity launched Computer for Max subscribers (with Pro and Enterprise rollouts following in March), featuring:
    • multi-model orchestration across 19 frontier models with Claude Opus 4.6 as the default orchestrator
    • more than 400 managed application integrations
    • configurable credit spending caps
    • 10,000 monthly credits on Max, paired with a limited-time 20,000-to-35,000-credit launch bonus
  • March 11, 2026: At its inaugural Ask 2026 developer conference, Perplexity unveiled Personal Computer for desktop workflows, Computer for Enterprise (adding Slack integration, Snowflake connectors, and 20 orchestrated models), Comet Enterprise, and the expanded developer API platform, while establishing the Secure Intelligence Institute for agent security research.
  • June 18, 2026: Perplexity introduced Brain, a persistent project-memory system for Computer that retains prior work trajectories, verified sources, and user corrections across sessions.
  • July 2026: Perplexity open-sourced Numbat, its endpoint detection and policy-enforcement suite for developer and agent harnesses, alongside Bumblebee for software supply-chain scanning, and launched Personal Computer on Windows (July 28).
  • August 25, 2026: Perplexity and NVIDIA launched Portable Computer, a local-first edition of Computer for DGX Spark and RTX-equipped Linux workstations where tasks run on device by default at zero credit cost.
  • September 2026: Perplexity shipped a rapid sequence of Computer platform upgrades:
    • September 1: Hybrid Compute on Mac, splitting tasks between cloud frontier models and local Apple Silicon checkpoints behind an on-device privacy gate
    • September 14: Portable Computer support for Windows PCs equipped with Nvidia RTX GPUs (24 GB+ VRAM)
    • September 15: HP began preloading the Perplexity Windows app on new PCs, and Crusoe signed a multi-year Nvidia GB300 compute agreement with Perplexity
    • September 17: granular effort controls ranging from Light to Ultra
    • September 21: Perplexity detailed its reinforcement and post-training pipeline for PPLX 27B using real-world Computer trajectories to cut tool errors and user corrections
    • September 23: Perplexity published its internal red-team evaluation of the SPACE microVM sandbox ("Escaping SPACE, Part I")
    • September 24: Portable Computer expanded to Windows PCs powered by AMD Ryzen AI Max processors
    • September 27 to 28: Perplexity completed the transition from legacy Sonar chat completions to the unified Perplexity Agent API and launched Reusable Agents with versioned Profiles, Skills, and managed connectors
    • September 29: Automations replaced Scheduled Tasks, adding live event triggers from Slack, Gmail, Outlook, Linear, and GitHub alongside persistent memory of prior runs, and Perplexity published its defense-in-depth agent security architecture
    • Also during September: Computer launched as a native Microsoft Teams application, and Personal Computer opened to all paid Mac subscribers
  • October 1, 2026: Perplexity and American Express released 10 specialized business Skills for Computer covering cash-flow forecasting, marketing audits, candidate sourcing, and vendor diligence.

Additional Viewing

Load video
YouTube

In "What is Perplexity Computer?" (published February 27, 2026, about 38 minutes), entrepreneur and Startup Ideas Podcast host Greg Isenberg conducts a live, unscripted hands-on test of Perplexity Computer immediately after its Max launch. Isenberg walks through connecting Gmail via OAuth, spawning parallel subagents to research 30 to 50 companies and venture capital firms, setting up a daily 8:00 a.m. competitive-intelligence monitor, and generating a multi-page financial diligence PDF with charts on Shopify, while also highlighting practical supervision lessons when an agent sends outreach emails or asks for approval before high-credit batch runs. The video directly illustrates the parallel subagent orchestration, connector workflows, and credit-consumption trade-offs explained in The Mechanism, The Physics, and Known Limitations. What is Perplexity Computer? | Greg Isenberg

Load video
YouTube

"Perplexity Computer: The Super Agent Playbook (5 Real Workflows)," from HubSpot's Marketing Against the Grain with Kipp Bodnar and Kieran Flanagan (published March 5, 2026, about 25 minutes), demonstrates five production marketing and strategy workflows built inside Perplexity Computer. The hosts show how Computer spawns parallel batches of Claude Sonnet 4.6 subagents to scrape and analyze 100 bestselling business book covers into a reusable `SKILL.md` file, builds an interactive website tracking live Polymarket political and economic odds, reverse-engineers the growth loops of 15 fast-growing B2B companies, and audits HubSpot's own product pages using a custom ranking algorithm. It provides a concrete look at how teams build and iterate on the modular Skills and disposable web dashboards described in Why This Matters and Getting Started. Perplexity Computer: The Super Agent Playbook | Marketing Against the Grain

Load video
YouTube

"Webinar Replay: Securing AI with Perplexity" (published August 13, 2026, about 48 minutes) is an official technical session featuring Perplexity Chief Security Officer Kyle and Customer Success Lead Scott examining how Perplexity secures Computer, Comet, and enterprise deployments. Kyle explains the architecture of the SPACE Firecracker sandbox, how the network gateway credential broker injects OAuth tokens outside the agent's context window, how enterprise administrators enforce human-in-the-loop email approvals and role-based access control (RBAC) via Okta and Entra ID, and how Perplexity's open-source BrowseSafe, Numbat, and Bumblebee tools defend against prompt injection and supply-chain attacks at machine speed. The presentation is the definitive primary source for the security engineering, Brain memory controls, zero data retention, and Bring Your Own Key (BYOK) encryption covered across this section. Webinar Replay: Securing AI with Perplexity

Perplexity Personal Computer

Personal Computer is the desktop edition of Perplexity Computer, the multi-model AI agent that Perplexity launched in February 2026. Where Perplexity Computer runs entirely in the cloud, Personal Computer is built into the Perplexity desktop app for Mac and Windows and can also reach the user's own machine: approved folders and files, native applications, the Comet browser, and the desktop itself. Perplexity announced it on March 11, 2026, at Ask 2026, the company's first developer conference, began rolling it out to Mac users on April 16, 2026, and released a Windows version on July 28, 2026. The company describes the product as "a digital proxy for you," and its own help center calls Personal Computer the superset: it does everything Computer does, plus local work.

The idea behind Personal Computer is that an agent becomes far more useful when it can act where a person's work actually lives. A cloud agent can research the web and use connected services, but it cannot open a spreadsheet sitting in a Downloads folder, rename files in a project directory, or read a to-do list in Apple Notes. Personal Computer closes that gap by combining Perplexity's cloud orchestration, which splits a goal into subtasks and assigns them to specialized subagents running on more than 20 frontier models, with a local client that can carry out steps on the user's computer. Unlike traditional robotic process automation, which replays fixed scripts, the agent decides how to complete each step from the goal it was given.

A common misunderstanding is that Personal Computer is a piece of hardware. Early coverage focused on the Mac mini because Perplexity recommends running the software on an always-on Mac mini, and Apple executives later pointed to agent tools such as Personal Computer when discussing strong Mac mini demand, but Perplexity does not sell a device. Personal Computer is software that runs on any supported Mac (macOS 15 Sequoia or later) or Windows PC (Windows 10 version 1809 or later, or Windows 11), and a dedicated machine simply lets the agent keep working around the clock.


Personal Computer moves Perplexity's cloud agent onto the user's own desktop, trading the clean isolation of a remote sandbox for direct access to local files and apps, which makes the permissions a user grants the most important setting in the product.


How Personal Computer Works

Personal Computer is a hybrid system. Planning, web search, and most model reasoning happen on Perplexity's servers, where Computer runs tasks in an isolated sandbox and orchestrates tools, files, more than 400 app connectors, and personal context. The desktop app provides the local half: it reads, writes, and searches files in folders the user connects, works with native applications on the Mac, and can drive the Comet browser to operate web-based tools that have no direct connector, such as filling out forms or booking appointments. A single request can therefore combine a local CSV file, a web source, an email in Outlook, and a record in a connected service like Salesforce, Snowflake, or HubSpot, with the finished result saved back to disk.

The orchestration layer is what distinguishes Perplexity's approach from single-model assistants. Computer treats a request as an objective rather than an instruction, breaks it into parts, and routes each part to the model best suited for it, a design Perplexity summarizes with the line that a traditional operating system takes instructions while an AI operating system takes objectives. Since June 18, 2026, Computer has also drawn on Brain, a research-preview memory system for Max and Enterprise Max subscribers that records what the agent has done, which sources helped, and which corrections were applied, so that each new task starts with context from earlier work. Perplexity reports that Brain improved answer correctness by 25% and cut the cost of context-heavy tasks by 13% in its internal tests, figures that have not been independently verified.

People interact with Personal Computer in several ways. On a Mac, pressing both Command keys summons a floating task bar from any app, and the agent can see the active app and suggest quick actions; it also accepts spoken requests through voice mode. Tasks can be started or approved from an iPhone or another device signed in to the same account, so a user can ask a Mac mini at home to finish work while away, and Perplexity secures Mac mini sessions with two-factor authentication. On Windows, a Remote access permission allows desktop sessions to be started from the Perplexity web and mobile apps, so a task begun on a phone can be completed on the PC.

Hybrid Compute and Local Models

On September 1, 2026, Perplexity added Hybrid Compute to the Mac app, a mode that splits each Computer task between frontier models in the cloud and a smaller local model running on the Mac. The cloud keeps planning, web search, and heavy reasoning, while the local model handles private files, sensitive information, and on-device actions. A privacy gate on the Mac decides what may leave the machine: it can mask sensitive details such as names, addresses, and account numbers, keep a step local, refuse an action, or ask the user for consent before protected file information reaches the cloud. Work handled by a local model consumes no Computer credits, so Hybrid Compute can lower costs as well as improve privacy, at the price of some capability on the steps kept local.

Hybrid Compute launched with three local model options: Google's Gemma 4 E4B, Alibaba's Qwen3.6 35B-A3B (a mixture-of-experts model that activates only a few billion of its 35 billion parameters at a time), and a Perplexity model post-trained for Computer. According to Perplexity's help center, it requires an Apple silicon Mac running macOS 15 or later with at least 24 GB of unified memory, and 32 GB gives the best results. The same help article describes a related on-device mode for Windows PCs with AMD Ryzen AI Max processors or NVIDIA RTX graphics cards with at least 24 GB of GPU memory, where users download a model (about 20 GB of disk space) from Menu > Settings > Local Inference and select it from the composer's model menu.

Personal Computer, Perplexity Computer, and Portable Computer

Perplexity now sells three closely related agents that differ mainly in where the work runs and how local data is governed. Perplexity Computer is the cloud original, Personal Computer extends it to the user's desktop, and Portable Computer, launched with NVIDIA on August 25, 2026, runs the whole agent stack locally on hardware such as the NVIDIA DGX Spark and RTX-equipped Linux or Windows machines, reaching the cloud only with the user's permission. Choosing among them depends on whether an organization prioritizes zero local setup, direct access to everyday desktop files and native apps, or strict on-device isolation where neither model weights nor file contents touch an external server. The table below summarizes the operational differences across all three editions.

Feature Perplexity Computer Personal Computer Portable Computer
Where it runs Perplexity's cloud sandbox Hybrid: cloud orchestration plus a local client on Mac or Windows Entirely on the user's hardware, escalating to the cloud only with permission
Local files and apps No (connectors and uploads only) Yes: approved folders; native app control on Mac Yes
Models 20+ frontier models in the cloud Same cloud models, plus optional local models through Hybrid Compute (Mac) or local inference (supported Windows PCs) Local open-weight models such as PPLX 27B and Qwen 27B
Always-on use Runs in the cloud while you are away Runs 24/7 if the host computer stays on (a Mac mini is recommended) Runs on a dedicated local machine
Typical buyer Anyone on a paid plan who wants a cloud "digital worker" Knowledge workers whose files and tools live on their own computer Privacy-focused users and teams with capable local GPUs
Cost basis Computer credits Computer credits for cloud steps; local steps are free No token costs for local work

Capabilities

Personal Computer is designed for multi-step knowledge work that mixes local material with online information. Perplexity's own examples include reading a to-do list in Notes or Notepad and completing each item across local files, email, connected apps, and the web; sorting a cluttered Downloads folder into project folders; comparing two documents and summarizing their differences; analyzing a CSV file and cross-checking it against a web source; and building a live dashboard instead of a static report. Because outputs are saved as ordinary files, the agent can produce spreadsheets, documents, and slide decks that a user can open, edit, and share without leaving their usual tools.

The Windows release leaned into Microsoft Office workflows. Perplexity's launch examples include refreshing a local PowerPoint pitch deck with current market and competitor data, building a quarterly board update for finance from a OneDrive folder and a local Excel forecast, redlining a purchase agreement for a legal team using terms from a local Excel cap table, and preparing a regional sales review from SharePoint and an Excel export. These build on Computer in Microsoft 365, which arrived in May 2026 as a side panel in Excel, Word, PowerPoint, and Outlook, and on Computer in Microsoft Teams. Perplexity also claims that Computer has performed more than $9.4 billion in labor-equivalent work for its users, a company figure that has not been independently audited.

The two desktop versions are not yet identical. According to Perplexity's help center as of late September 2026, the Mac version can control installed apps such as Mail, Finder, Slack, Messages, and Notes, while installed-app control is not yet available on Windows, where Personal Computer works with local files, connectors, and the web. On both platforms, the agent can only see the folders and services the user connects, and the quality of its work depends heavily on how clearly a task is described, a reminder that prompt engineering still matters even when the agent plans its own steps.

Capability Mac Windows
Local file access (read, write, search approved folders) Yes Yes
Control of installed apps Yes (Mail, Finder, Slack, Messages, Notes, and others) Not yet available
Comet browser control for web tools without connectors Yes Yes
Voice requests Yes Yes
Start or approve tasks from a phone Yes Yes (with Remote access enabled)
On-device models Hybrid Compute on Apple silicon with 24 GB+ memory Local inference on supported AMD and NVIDIA hardware
System requirement macOS 15 Sequoia or later Windows 10 version 1809 or later, or Windows 11

Security, Privacy, and Control

Giving an agent access to a personal computer raises the stakes compared with a cloud-only assistant, and Perplexity has built several layers of control around it. Computer executes actions in an isolated sandbox, files it creates are kept in a secure sandbox, and on-device authorization protects sensitive steps. Personal Computer alerts the user before sensitive actions such as sending an email or deleting a file, every session keeps an audit trail so actions can be reviewed, and Perplexity describes the Mac version's actions as auditable and reversible, with a kill switch that stops the agent immediately. Perplexity also states that it does not train on company data in its Enterprise plans, and administrators can decide whether local inference is allowed and, on Mac, which kinds of data must stay on the device.

The operating-system permissions that make the product useful are also its biggest risk. On a Mac, the app asks for Accessibility (to identify windows and apps), Screen Recording (for visual context), Full Disk Access (to search files across the Mac), and Microphone (for voice), and controlling apps can trigger additional Automation prompts; on Windows, users enable Local file system access and Desktop tools. An agent that reads web pages and email can also be targeted by hidden instructions in that content, a prompt injection attack, and Perplexity's Comet browser drew public attention in 2025 when security researchers demonstrated such an attack, which they called "CometJacking." Sensible practice follows ordinary risk management: connect only the folders a task needs, prefer specific folders over Full Disk Access, keep approvals on for anything that sends, deletes, or buys, and keep a person in the loop for review, a pattern known as human-in-the-loop oversight.

Recent Developments

  • February 25, 2026: Perplexity launches Perplexity Computer, a cloud "digital worker" that orchestrates 19 models, initially for Max subscribers; it reaches Pro subscribers in March.
  • March 11, 2026: At Ask 2026, Perplexity announces Personal Computer, Computer for Enterprise, Comet Enterprise, and a developer API platform, and opens a Personal Computer waitlist.
  • April 16, 2026: Personal Computer for Mac begins rolling out to Max subscribers and waitlist members.
  • May 2026: A new Perplexity Mac app makes Personal Computer available to all paid Mac users (May 7), and Computer arrives in Microsoft 365 apps and Microsoft Teams.
  • June 2026: Perplexity announces the Windows version (June 3) and launches Brain, a persistent memory system for Computer (June 18).
  • July 28, 2026: Personal Computer for Windows launches for Pro, Max, and Enterprise subscribers on Windows 10 and 11.
  • August 25, 2026: Perplexity and NVIDIA launch Portable Computer, a fully local version of Computer for DGX Spark and RTX-equipped Linux machines.
  • September 1, 2026: Hybrid Compute brings local models and an on-device privacy gate to Personal Computer on Apple silicon Macs.

These steps show a clear strategy: Perplexity is turning its answer engine into a platform for agents that work across the cloud, the desktop, and local hardware, competing directly with Microsoft Copilot, Claude Cowork, OpenAI's ChatGPT Work and dots, and Google's Gemini Spark. Perplexity's differentiators are model choice, grounding in cited search results, and an unusually flexible deployment story that ranges from fully cloud to fully local. The open questions are cost predictability under a credit system, reliability on long unattended tasks, and whether users will be comfortable granting an agent the broad operating-system permissions it needs. For side-by-side comparisons, see the Ready-to-Use AI Agent Comparison.

Getting Started with Personal Computer

A beginner can install Personal Computer and complete a first useful task in a single sitting. The safest approach is to start with a small, low-risk job in a folder created just for testing, connect only that folder, and keep approvals on for any action that sends, deletes, or changes something outside the folder. The steps below follow Perplexity's help center as of late October 2026.

  1. Check your plan and computer. You need a paid Perplexity plan (Pro, Max, or Enterprise) and either a Mac running macOS 15 Sequoia or later or a PC running Windows 10 version 1809 or later or Windows 11. Organization-managed accounts may need an administrator to allow desktop tools.
  2. Install the app. On Mac, download the app from perplexity.ai/personal-computer, open the disk image (DMG), and drag Perplexity into Applications. On Windows, install the Perplexity app from the Microsoft Store or Perplexity's download page.
  3. Sign in and grant permissions. On Mac, open Perplexity > Settings… > Permissions and approve Accessibility; approve Screen Recording, Full Disk Access, and Microphone only when a feature needs them. On Windows, open Settings > Permissions, select Enable in the Local file system access row, and turn on Desktop tools (app version 26.8.1 or later).
  4. Connect a folder. In Perplexity Computer, open Folders, select Choose folder, and pick only the folder your task needs. You can add different folders for different tasks and remove them later.
  5. Connect apps and devices (optional). Connect the apps you want the agent to use through connectors, install Comet for browser tasks, and sign in on your phone with the same account to pair it; paired devices appear under Settings > Devices.
  6. Give a clear objective. On Mac, press both Command keys or open the app; describe the outcome, the sources it may use, what it must not do, and when to stop for your review.
  7. Review the work. Check the files it creates, the actions it took, and the credits used before relying on the result or expanding its access.

Working Example: Meeting Notes to an Action Tracker

This example uses three small text files, so it costs few credits, touches nothing important, and produces a result that can be checked by hand. First, create a folder named PC-Test (for example, in Documents) and save the three notes below in it as plain-text files with the names shown. Then connect only the PC-Test folder in step 4 above and give Personal Computer the objective in the second block. 

notes-2026-10-05.txt
Project Falcon kickoff. Attendees: Ana, Ben, Chloe.
Action: Ana to draft the project charter by Oct 12.
Action: Ben to request budget approval by Oct 9.
Decision: Weekly check-in every Tuesday.

notes-2026-10-13.txt
Project Falcon check-in. Attendees: Ana, Ben, Chloe.
Charter draft received from Ana.
Action: Chloe to collect vendor quotes by Oct 20.
Action: Ben to schedule a demo with the vendor shortlist by Oct 23.
Risk: Budget approval still pending.

notes-2026-10-20.txt
Project Falcon check-in. Attendees: Ana, Chloe.
Action: Ana to update the charter with vendor costs by Oct 27.
Action: Chloe to share the quotes with the team by Oct 21.
Ben absent; budget approval status unknown.

Use only the files in the PC-Test folder. Do not browse the web.

1. Read the three meeting notes.
2. Create an Excel workbook named falcon-actions.xlsx in a new subfolder
   called Output. On the first sheet, list one row per action item with:
   owner, task, due date, source file, and status (Done only if a later
   note says it was completed; otherwise Open).
3. On a second sheet, list the decisions and risks.
4. Draft, but do not send, a short follow-up email to Ana, Ben, and Chloe
   that summarizes open items and flags anything overdue as of Oct 21, 2026.

Do not move, rename, or delete any existing file. Show me the results
before doing anything else.

The result is easy to verify. The tracker should contain exactly six action items; Ana's charter draft should be marked Done because the October 13 notes confirm it was received; Ben's budget approval (due October 9) should be flagged as overdue in both the tracker and the draft email; and the decisions sheet should list the weekly Tuesday check-in, with the pending budget approval as the main risk. The three original notes should be untouched, the new workbook should sit in PC-Test/Output, and no email should have been sent. If anything is wrong, say so in the same conversation, for example "Ben's budget approval is overdue; update the status and the email," and check how many credits the run used so you can estimate the cost of larger jobs.

Once the example works, the same pattern scales to real work. You might connect a project folder and ask Personal Computer to turn each new set of meeting notes into tracker updates, or pair the agent with a connector such as Outlook or Slack and let it draft, but not send, weekly status summaries. On a Mac mini that stays on, a recurring version of the task can run while you are away, which is exactly the always-on use Perplexity designed the product for; just keep each expansion matched with an approval rule so the agent's growing reach never outpaces your oversight.

Additional Viewing

Load video
YouTube

Personal Computer, by Perplexity is the company's official announcement video, published on March 11, 2026, the day Personal Computer was unveiled at Ask 2026. It presents Perplexity's framing that a traditional operating system takes instructions while an AI operating system takes objectives, and it shows the agent working across local apps and files on an always-on Mac mini. The video is the best primary source for the product vision described in the Introduction/Overview and How Personal Computer Works sections. Watch on YouTube

Load video
YouTube

Perplexity Personal Computer - Your Mac Just Became an AI Agent is a 12-minute beginner tutorial from Teacher's Tech, published April 19, 2026, just after the Mac rollout began. The presenter explains what the agent does, its pricing, and how to install it, then demonstrates drafting an email, creating files from a prompt, and turning emailed or photographed receipts into a spreadsheet with a pie chart before weighing the product's limitations and cost. It complements the Getting Started section and the meeting-notes working example, which follow the same local-files-to-spreadsheet pattern. Watch on YouTube

Load video
YouTube

Perplexity Computer for Beginners: Everything You Need to Know is a roughly 16-minute walkthrough from Teacher's Tech, published March 4, 2026, covering the cloud agent on which Personal Computer is built. It explains how Computer orchestrates many models, how pricing and access work, and demonstrates five prompts, including a deep research report, an interactive dashboard, an app build, and an automated weekly briefing. Watching it first makes the Personal Computer, Perplexity Computer, and Portable Computer comparison easier to follow, because every Personal Computer task still relies on this orchestration layer. Watch on YouTube

References

Perplexity Comet Assistant

The Comet Assistant is the AI agent built into Comet, the Chromium-based web browser from Perplexity. It lives in a sidecar panel beside every web page, sees what the user is looking at, and can answer questions about the page, summarize it, compare it with other open tabs, and carry out web tasks such as filling out forms, comparing prices across sites, drafting emails, or building a shopping cart. Perplexity markets Comet as a browser that acts as a personal assistant, and the Comet Assistant is the part that does the work: instead of copying text into a separate chatbot, the user asks in place, and the assistant uses the page, the tabs, and, with permission, connected accounts as its context. Perplexity's own retrieval-grounded answer engine also replaces the usual list of search results, so cited answers appear directly in the browser.

Comet launched on July 9, 2025, for Windows and macOS, initially only for subscribers to Perplexity's $200-a-month Max plan and a waitlist. Perplexity made the browser and its sidecar assistant free worldwide on October 2, 2025, released an Android version on November 20, 2025, and brought Comet to the iPhone on March 18, 2026. On November 6, 2025, Perplexity shipped a rebuilt Comet Assistant that works across several tabs in parallel, takes on longer jobs, and asks permission before acting, and its internal tests showed the new version performing 23% better than its predecessor on multi-step tasks.

The assistant now operates in tiers. Free users get the sidecar and a mode for reading and interactive tasks that they watch and can interrupt, while paid Pro and Max subscribers can ask Comet to control the browser on their behalf, hand larger jobs to Perplexity Computer to run in the background, and use Background Assistants and an Email Assistant on the Max plan. Comet Enterprise, launched in March 2026, packages the same assistant with administrative controls, audit logs, and device-management deployment for companies.

The Comet Assistant differs from a chatbot because it acts inside the user's own logged-in browsing session rather than in a separate window, and it differs from cloud agents such as Perplexity Computer or Grok Bot because its default workspace is the browser on the user's own device. That design makes it convenient and fast for everyday web work, but it also exposes the assistant to everything it reads on the web, which is why prompt injection, permissions, and the legal status of agents acting on someone's behalf have shaped its development as much as any new feature. Understanding those trade-offs is the key to using it well, because the same access that makes it useful is what an attacker or a mistaken instruction would exploit.


The Comet Assistant turns the browser itself into the agent: instead of visiting the web and then asking an AI about it, you ask the web to do the work, inside your own tabs and logged-in accounts.


Mechanisms

The Mechanism

The assistant opens in a sidecar beside the current page with a keyboard shortcut (Alt+A on Windows, Option+A on Mac) or a toolbar button, and it can also be used by voice. By default its context is the page the user is viewing and any text they select. Users can pull in other open tabs by typing @ followed by a tab name, and they can explicitly share history or other material when a request needs it. Perplexity's March 2026 privacy documentation says the assistant does not, by default, upload the full browsing history, the full list of open tabs, cookies or site data, passwords and autofill data, local files, or text typed into websites, and that page, tab, or history context sent with a request can be retained for up to 30 days.

When the assistant acts on a page, the browser captures what is on screen and sends it to Perplexity's servers, which return navigation instructions that the browser carries out on the user's own device. This split mattered in court: in August 2026 the Ninth Circuit relied on it to conclude that the user, not Perplexity, is the one accessing a website when Comet shops on the user's behalf. Since the November 2025 rebuild, the assistant asks for approval before taking actions in the browser, such as clicking a button or entering data, and remembers the user's choice for the rest of that task. The work is organized into several levels of autonomy:

  • Search mode: the free default for answering questions, summarizing pages, PDFs, and videos, comparing tabs, and handling interactive tasks while the user watches and can step in
  • Control browser: a mode for Pro and Max subscribers in which Comet navigates the web on the user's behalf; Perplexity made Anthropic's Claude Opus 4.6 the default browser-agent model for Max subscribers when it introduced the mode
  • Work mode: a credit-metered option that hands a task to Perplexity Computer, which can spin up a browser agent inside Comet and reach logged-in sites with the user's permission, without connectors or MCP servers
  • Background Assistants: a Max feature, introduced in October 2025, that runs several assistants at once from a dashboard and can repeat tasks on recurring triggers
  • Email Assistant: a Max feature that drafts replies in the user's tone, organizes the inbox, and schedules meetings through connected Gmail and Google Calendar accounts

Several smaller features make the assistant easier to direct. Shortcuts save reusable prompts that can chain steps and reference tabs or files, Voice Mode (rebuilt on OpenAI's GPT Realtime 1.5) lets users ask about what is on screen and navigate by voice while switching tabs, and connections to Gmail and Calendar are strictly opt-in and revocable at any time. Because Comet is built on Chromium, it supports most Chrome extensions on the desktop, imports bookmarks and settings from Chrome, and accepts more than 500 Chromium management policies in enterprise deployments.

The Physics

An assistant that reads arbitrary web pages and acts with the user's own logins runs into hard limits of trust, reliability, and permission. Comet became one of the most studied examples of these limits because security researchers examined it closely in its first months. The main constraints are:

  • Indirect prompt injection: Brave's security team showed in August 2025 that instructions hidden in a Reddit comment could hijack the assistant when a user simply asked it to summarize the page, and in October 2025 it found that text invisible to people but readable in screenshots could do the same. LayerX's "CometJacking" research showed that a single crafted link could make the assistant pull data from connected services and send it out, and Guardio's "Scamlexity" tests tricked Comet into buying from a fake store and entering credentials on a phishing page.
  • Imperfect detection: Perplexity responded with layered defenses, including real-time classifiers that scan page content before the agent acts and its open BrowseSafe detection model. Perplexity reported a 91% detection rate on its own BrowseSafe-Bench benchmark, which still means roughly one attack in ten gets past the classifier, and detection dropped to 76% for multilingual attacks.
  • Reliability: Early hands-on testing found the agent helpful for simple tasks but prone to falling apart on complex ones, and independent reviewers have reported failure rates as high as a third on delegated multi-step jobs.
  • Platform permission: Websites can object to agents acting on users' behalf. Amazon sued Perplexity in November 2025 and won a preliminary injunction in March 2026 before the Ninth Circuit vacated it in August.
  • Identification: Interactive Comet sessions load pages as ordinary Chromium traffic from the user's own device and do not announce themselves the way Perplexity's declared crawlers do, which makes them hard for sites to recognize and is one reason some retailers have resisted them.
  • Capacity: Agentic use is metered. Perplexity publishes limits of 80 monthly browser-agent queries on Enterprise Pro and 800 on Enterprise Max, and consumer Work mode draws on Computer credits.

These constraints explain how careful users work with the assistant. From a cybersecurity standpoint, reading and summarizing tasks carry little risk, while anything that touches email, money, or accounts deserves narrow permissions, allow-once approvals, and close supervision. Security researcher Simon Willison argued after Brave's disclosure that trusted instructions and untrusted content still end up in the same stream of tokens for a large language model, so defenses reduce the risk of prompt injection without eliminating it, a view that Perplexity's own description of BrowseSafe as one layer among several reflects.

Why This Matters

The Comet Assistant is Perplexity's attempt to own the starting point of people's time online. CEO Aravind Srinivas has argued that becoming the default browser can deliver "infinite retention," and the assistant is what gives users a reason to switch from Chrome or Safari. Its significance reaches beyond one product:

  • A legal precedent for agents: The Ninth Circuit's August 4, 2026, ruling in Amazon's case held that the user, not Perplexity, accesses a site when Comet acts at the user's direction, and the court warned that an injunction would impair consumer choice and limit a nascent technology. Amazon's trademark and state-law claims continue, and the court called its holding narrow.
  • Open security research: Perplexity released BrowseSafe as an open-weight detection model along with BrowseSafe-Bench, a public benchmark of 14,719 examples that mimic production web pages, giving other agent builders a ready-made layer of protection.
  • Publisher economics: Comet Plus, announced in August 2025 at $5 a month and included in Pro and Max, promised participating publishers 80% of its revenue from an initial $42.5 million pool, paid for human visits, search citations, and agent actions.
  • Distribution beyond Comet: Perplexity turned what it learned building Comet into APIs that power Samsung's Browsing Assist, a conversational assistant built into Samsung Browser on Galaxy phones and Windows PCs.
  • A surviving standalone AI browser: OpenAI shut down its ChatGPT Atlas browser on August 9, 2026, after folding it into its desktop apps, leaving Comet as one of the few actively developed browsers built around an assistant from the start.
  • Enterprise adoption: Comet Enterprise added MDM deployment, telemetry and audit logs, CrowdStrike Falcon protection against phishing and malware, and a promise that enterprise data is not used for training, with early customers including Fortune, AWS, AlixPartners, and Bessemer Venture Partners.

The assistant competes with a growing set of personal agents, including Meta Muse, Grok Bot, OpenAI's dots, Google's Gemini Spark with Chrome auto browse, and Anthropic's Claude Cowork and Claude in Chrome. Most of those run in cloud computers that the vendor manages, while Comet keeps the user's own browser at the center, which gives it direct access to existing logins but also places it squarely on the front line of web security. Side-by-side reviews are collected in the Ready-to-Use AI Agent Comparison.

Building Blocks

Component What It Does Notes
Comet browser A Chromium-based browser with Perplexity as the default answer engine, built-in ad blocking, and support for most Chrome extensions on desktop. Available on Windows 10 and later, macOS Big Sur and later, Android 12 and later, and iOS 18 and later. The iOS version lacks extensions.
Assistant sidecar A panel beside every page that reads the current page and selected text and answers or acts in place. Opened with Alt+A (Windows) or Option+A (Mac); the @tab syntax brings other open tabs into a request.
Search mode Free mode for questions, summaries of pages, PDFs, and videos, cross-tab comparisons, and supervised interactive tasks. The user watches and can step in at any time.
Control browser Lets Comet navigate and act on the web on the user's behalf. Pro and Max; Max uses Claude Opus 4.6 as the default browser-agent model.
Work mode and Perplexity Computer Hands longer jobs to Perplexity Computer, which can launch a browser agent inside Comet that reaches logged-in sites with permission. Credit-metered; Max includes 10,000 credits a month. Since October 2026, Comet is the default browser for Computer's browsing.
Background Assistants Run several tasks at once from a dashboard, including recurring tasks. Max plan.
Email Assistant Drafts replies in the user's tone, organizes the inbox, and schedules meetings through Gmail and Google Calendar. Max plan; account connections are opt-in and revocable.
Shortcuts Saved, reusable prompts that can combine steps and reference tabs or files. Useful for routines such as daily briefings or recurring comparisons.
Voice Mode Lets users ask about what is on screen, navigate by voice, and keep the conversation as they switch tabs. Rebuilt on OpenAI's GPT Realtime 1.5.
Permissions and privacy controls Asks before acting in the browser and remembers the choice for the task; users can block the assistant on specific sites and delete threads. By default the assistant does not upload full history, tab lists, cookies, passwords, local files, or typed text.
Prompt injection defenses Real-time classifiers and the open BrowseSafe model scan page content for hidden instructions before the agent acts. BrowseSafe reports a 91% detection rate on Perplexity's own benchmark.
Comet Enterprise Managed deployment with admin controls, telemetry, audit logs, 500+ Chromium policies, and CrowdStrike Falcon integration. Launched March 2026; Enterprise Pro and Max include 80 and 800 monthly browser-agent queries.

Getting Started

Comet installs like any other browser, and the assistant is available as soon as it opens. The choices that matter most concern what the assistant may see and do, so it is worth setting permissions before handing it anything sensitive. The steps below follow Perplexity's setup guidance and independent walkthroughs:

  1. Download Comet: Get the installer from Perplexity's Comet page for Windows or macOS, or install the app from Google Play or the App Store. Comet is free; basic browsing and some AI features work even without a Perplexity account.
  2. Import your browser data: On first launch, import bookmarks, settings, and extensions from Chrome or another browser so Comet feels familiar.
  3. Sign in: Sign in with a Perplexity account to sync across devices and use the assistant fully. A Pro or Max subscription unlocks Control browser, Work mode credits, and, on Max, Background Assistants and the Email Assistant.
  4. Open the assistant: Press Alt+A (Windows) or Option+A (Mac), or tap the assistant button on mobile. Option+Shift+V starts Voice Mode on a Mac.
  5. Start with the current page: Ask the assistant to summarize, explain, or extract information from the page you are reading, then use @ followed by a tab name to bring other tabs into the question.
  6. Review privacy settings: Block the assistant on sensitive sites such as banking or health portals, and delete temporary threads when a task is finished.
  7. Connect accounts only when needed: If you want help with email or scheduling, grant Gmail or Calendar access through the Perplexity connector. Access is opt-in and can be revoked at any time in settings.
  8. Approve actions deliberately: When the assistant asks to click, type, or submit, choose to allow once until you trust a workflow, and watch the first few runs.
  9. Hand off a first task: Describe the outcome, the sites to use, and any limits, such as a budget or dates. Comparison and research tasks are good starting points before anything that buys or sends.
  10. Save repeat requests as Shortcuts: Turn prompts you use often, such as a morning briefing or a weekly price check, into Shortcuts.
  11. Deploy for a team, if applicable: Administrators can roll out Comet Enterprise through existing MDM tools, apply existing Chrome policies by swapping the Chrome identifier for Comet's, and monitor usage through audit logs.

A few habits reduce risk considerably. Keep a human in the loop for purchases, sending email, and changing account settings, avoid giving the assistant access to everything at once, and be suspicious of pages, emails, or links from unknown sources while the assistant is active, since those are the usual vehicles for prompt injection. Treat the assistant's output as a draft to verify, especially for prices, dates, and anything it fills into a form.

Examples

Perplexity, reviewers, and early users have described a wide range of tasks for the Comet Assistant. The examples below are drawn from Perplexity's announcements, independent reviews, and published demonstrations rather than invented scenarios. Most fall into a few groups, and the safest and most reliable ones involve reading and organizing information, while tasks that act on accounts benefit from closer supervision.

Research and Reading

  • Summaries: Summarize a long article, a PDF, or a YouTube video without leaving the page, then ask follow-up questions or request an explanation of unfamiliar terms.
  • Cross-tab comparison: Build a table comparing products or services across open tabs, for example project tools such as Notion, Asana, Monday, and ClickUp on price for a team of ten, integrations, and user sentiment.
  • Deep research: Run Perplexity's Deep Research from the browser to gather information from many sources into a cited summary.
  • Job and market research: Search LinkedIn for specific job roles, or analyze competitors' listings and pricing in a market.

Shopping and Travel

  • Travel deals: Perplexity's own example asks Comet to search a few major sites for the cheapest flight from San Francisco to New York next Monday.
  • Price comparison: Compare prices and shipping speeds for the same product across several stores, and assemble a shopping cart for review.
  • Bookings: Research hotels or reservations and fill in booking forms, stopping for the user's approval before submitting.

Email, Calendar, and Daily Planning

  • Inbox help: Summarize unread email, draft replies in the user's tone, and organize the inbox with the Email Assistant.
  • Calendar briefings: Review the day's calendar, brief the user on each meeting and its attendees, and suggest when to leave and how to get there by public transit.
  • Scheduling: Find times and set up meetings through a connected calendar.

Work and Productivity

  • Data gathering: Collect details from several websites or online portals and compile them into an organized summary or a Google Sheet, such as an attendance or expense tracker.
  • Form filling: Complete repetitive web forms, such as a vendor security questionnaire, from information the user supplies.
  • Screening and review: Review applicants or proposals on a marketplace such as Upwork against criteria the user sets.
  • Enterprise analysis: At Atrium Hospitality, an AI manager described using Comet to pull context from separate hotel brand systems and draft analysis before a morning revenue call.

Sample Requests

The table below shows illustrative requests built from documented capabilities, with the mode each typically uses and where the user should expect to approve. Comet accepts ordinary language, so stating the goal, the sources, and the limits matters more than exact wording. Approval points reflect the assistant's practice of asking before it clicks, types, or submits on a site.

Request Typical Mode Where You Approve
"Summarize this article and list the three claims I should verify." Search mode No approval needed
"Compare the pricing pages in my open tabs and make a table for a team of ten." Search mode with @tab No approval needed
"Search a few major sites and find me the cheapest flight from San Francisco to New York next Monday." Search mode or Control browser Before entering traveler or payment details
"Collect the contact details from these five vendor portals into a Google Sheet." Control browser or Work mode Before signing in or writing to the sheet
"Draft replies to today's unread emails from clients and flag anything urgent." Email Assistant Before sending any reply
"Every Monday, check these three stores for the price of this monitor and tell me if it drops." Background Assistant or Shortcut Before buying

Known Limitations

  • Prompt injection remains unsolved: Despite classifiers and BrowseSafe, Perplexity's own figures imply that some malicious pages will slip past detection, and researchers continue to publish new attack variants against agentic browsers.
  • Inconsistent results on complex tasks: Reviewers report that the assistant handles simple tasks well but can stall, misread dates, or give up on long, multi-site jobs, so its work needs checking.
  • Data leaves the device: Page, tab, or history context the user sends with a request goes to Perplexity's servers and can be retained for up to 30 days, and agentic actions send screenshots of the page to Perplexity for processing.
  • Paywalls on autonomy: The free tier covers reading and supervised tasks, while Control browser, Work mode, Background Assistants, and the Email Assistant require Pro or Max, and agentic use is metered by credits or monthly query limits.
  • Site resistance and legal uncertainty: Amazon's lawsuit continues on trademark and state-law grounds even after the Ninth Circuit vacated the injunction, and other sites may block agents that do not identify themselves.
  • Publisher disputes: Perplexity faces several publisher lawsuits, including one alleging that the Comet browser circumvents a newspaper's paywall, which could affect what the assistant may read.
  • Platform gaps: The iOS version does not support extensions, and some features reach desktop and Android before iPhone.

History

Recent Developments

  • February 2025: Perplexity announced Comet without a release date.
  • July 9, 2025: Comet launched for Windows and macOS for Max subscribers and a waitlist, with the Comet Assistant in a sidecar.
  • July 25 to August 20, 2025: Brave reported an indirect prompt injection flaw to Perplexity, Perplexity shipped fixes, and Brave published its findings, noting that the issue was not fully mitigated.
  • August 25, 2025: Perplexity announced Comet Plus, a $5 monthly subscription sharing 80% of revenue with publishers from an initial $42.5 million pool.
  • August 2025: LayerX discovered the "CometJacking" technique, in which a crafted link could make the assistant exfiltrate data from connected services.
  • October 2, 2025: Comet became free worldwide, Background Assistants arrived for paid subscribers, and Comet Plus launched with publishers including The Washington Post, Le Monde, and Condé Nast. A Max-only Email Assistant had been announced days earlier.
  • October 21 to 22, 2025: Brave disclosed screenshot-based prompt injection in Comet, and Perplexity's security team published "Mitigating Prompt Injection in Comet."
  • November 2025: Amazon sued Perplexity over Comet's shopping on Amazon. On November 6, Perplexity released the rebuilt Comet Assistant with multi-tab work, longer tasks, a new permission flow, and a 23% improvement in internal tests, and on November 20 Comet launched on Android.
  • December 2, 2025: Perplexity released BrowseSafe, an open prompt-injection detection model, and the BrowseSafe-Bench benchmark.
  • February 25, 2026: Perplexity launched Perplexity Computer, its multi-model agent for long-running work, which later gained the ability to take control of Comet.
  • March 9, 2026: A federal district court granted Amazon a preliminary injunction against Comet's shopping agent; the Ninth Circuit stayed it pending appeal.
  • March 18, 2026: Comet launched on iPhone, completing its set of platforms, and Perplexity introduced Comet Enterprise.
  • August 4, 2026: The Ninth Circuit vacated the injunction, ruling that users rather than Perplexity access Amazon when Comet acts at their direction. Amazon petitioned for rehearing by the full court on August 18, and the appeals court rejected that bid the following month.
  • August 9, 2026: OpenAI shut down its ChatGPT Atlas browser, leaving Comet as a leading standalone AI browser.
  • October 5, 2026: Comet became the default browser for Perplexity Computer's browsing work, alongside new scheduled and conditional Automations in Computer.
  • October 2026: Comet for iOS added layouts optimized for the foldable iPhone Duo, and the desktop browser reached version 156.

Further Resources

Additional Viewing

Load video
YouTube

Perplexity CEO Aravind Srinivas on Comet, AI Browser Wars, and Why He's Not Selling to Big Tech is a CNBC interview in which Deirdre Bosa talks with Srinivas about the launch of Comet as a browser designed to go beyond chatbots and handle tasks. It explains the strategy described in the Introduction/Overview and Why This Matters sections, including why Perplexity built its own browser rather than an extension. Watch on YouTube

Load video
YouTube

Perplexity CEO Aravind Srinivas On Comet, Search, & The Future Of AI is a wide-ranging conversation with Semafor technology editor Reed Albergotti about the new Comet browser, search, and where AI assistants are headed. It complements the competitive and business discussion in Why This Matters, from distribution to the economics of answer engines. Watch on YouTube

Load video
YouTube

Securing AI Agents: How to Prevent Hidden Prompt Injection Attacks is a technical explainer from IBM Technology on how agents that browse the web can be manipulated by instructions hidden in content and what defenses developers can apply. It gives general background for the prompt injection constraints in The Physics and the layered defenses described in the Building Blocks table. Watch on YouTube

Load video
YouTube

39C3 - Agentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents is a Chaos Communication Congress talk by security researcher Johann Rehberger, published by media.ccc.de in January 2026. It demonstrates end-to-end prompt injection exploits against computer-use and coding agents, including data exfiltration and command-and-control takeovers, and discusses vendor responses and mitigations, illustrating why the defenses in The Physics remain only partial. Watch on YouTube

References

Retrieved from "https://primo.ai/w/index.php?title=Perplexity&oldid=53528"