Overview
Perplexity is a search product that puts AI conversation and live web search in the same interface. Every answer retrieves the web in real time and lists its sources underneath, so each claim can be checked by opening the original page. Alongside its own Sonar line, it offers models from OpenAI, Anthropic, Google, Moonshot, Z.ai and xAI, with the selectable set depending on the plan.
Through the second half of 2026 it widened its focus from search into the Computer agent. Projects lets it run several tasks in parallel with persistent context, the Windows and macOS apps take over local files and connected applications, and in October the macOS app added a fully local mode where both the model and the agent run on the machine, so sensitive material never has to leave it.
Key Features
- Live retrieval with cited sources: Each answer searches the web first, synthesises several sources and lists the links underneath so every claim can be opened and checked.
- Pro Search deep research: Complex questions are broken into several rounds of retrieval; you can follow up and narrow the scope, compressing background research into one conversation.
- Model switching: Beyond its own Sonar line, models from OpenAI, Anthropic and Google are selectable by plan, so the same question can be put to another model in a click.
- Computer agent: It drives the browser and local apps through multi-step tasks, Projects runs several of them in parallel with persistent context, and work can be handed over by email.
- Local mode on desktop: The macOS app runs without a network connection, keeping both model and agent on the machine, which suits contracts and financial statements that should not be uploaded.
- File analysis and Collections: PDFs and images can be uploaded and answered against search results, and both results and conversations can be grouped into themed collections for later reference.
Use Cases
- Researchers and analysts tracking the latest news and live data
- Professionals writing reports or competitive analysis who must cite every source
- Students and researchers doing literature retrieval and background checks
- Users with piles of PDFs and screenshots who would rather ask than page through
- People who want research, tidying and price comparison handled by an agent
- Users handling contracts or financial statements that should not leave the machine
Pros
- Claims are checkable, since every answer carries source links that open the original
- Strong real-time coverage, not limited by a training data cutoff
- Pro Search genuinely saves time on complex questions, converging after a few follow-ups
- The free tier covers daily use with a set number of standard and Pro searches
- Desktop local mode keeps sensitive material on the machine
- Search extends into execution, with Computer acting on what it just found
Pricing
The free tier includes a limited number of standard searches and 5 Pro Searches per day. Pro costs $20 per month or $200 per year and unlocks unlimited Pro Search, model switching and file uploads; Team starts at $40 per person per month, and Enterprise is quoted on request. Local mode on desktop is available to Pro, Max and Enterprise subscribers. See the official site for current pricing and quotas.
Summary
Perplexity puts finding and verifying in one interface, with sources sitting next to the answer, which suits anyone who has to check claims repeatedly. Through late 2026 it extended into the Computer agent, and the macOS app can now run entirely offline. It fits daily search, competitive research and literature review, with the clearest payoff in report writing that needs a citation per claim.
Version History
- Computer skill marketplace adds open source /impeccable skill for layout, type and accessibility review (2026-10-24): Perplexity added the open source /impeccable skill by pbakaus to the Computer skill marketplace on October 24. It pulls interfaces away from the stock AI look by checking layout, typography, contrast and accessibility, then fixing what a designer would flag. Each user gets up to 5000 credits free through October 31.
- Agent API batch mode, decision model runs locally (2026-10-23): Perplexity's Agent API gained a Batch API supporting up to 50,000 requests per file with web_search and fetch_url available in batch mode; from 23 October its decision model can run entirely on a Mac with data staying on device.
- Perplexity Computer adds a fully local mode on macOS (2026-10-20): The macOS desktop app now offers a fully local mode where both the model and the agent run on the machine, with no network connection and no paid credits consumed. Local mode is driven by PPLX Pro, a post-trained version of Qwen 3.8, with 36GB of unified memory recommended; heavier work can switch to a Hybrid mode that calls cloud models.
- Perplexity Computer adds fast inference for GPT-6 Astra and SOL 6.1 plus an /anatomy skill (2026-10-14): On October 14, 2026, Perplexity brought fast inference versions of GPT-6 Astra and SOL 6.1 to Computer, along with an /anatomy skill and support for Polymarket tags. Perplexity notes that its sandbox platform handles millions of executions a day, routing each request to a suitable runtime.
- Open sources pplx-embed-v2-late multimodal embeddings (2026-10-08): Perplexity open sourced pplx-embed-v2-late, two late-interaction multi-vector embedding models for text and images at 9B and 0.6B that share a single embedding space, with weights on Hugging Face. The 9B indexes multimodal data while the 0.6B handles on-device queries, retrieving PDF pages without OCR. Reported scores are 92.4 percent on MADQA and 64 percent on BrowseComp+
- Computer email delegation (2026-09-30): Perplexity opened an email delegation entry point for Computer and made task runs free for a limited period, letting users hand a task to the agent by mail
- Perplexity's self-developed CobbleDB replaces AWS DynamoDB, saving up to one hundred million dollars annually. (2026-09-15): Perplexity CEO Aravind Srinivas announced the self-developed key-value database CobbleDB, replacing AWS DynamoDB for fast web content crawling, which can save up to one hundred million dollars per year after migration. The project's core infrastructure was completed in two months by two engineers and hundreds of continuously running Computer agents. Official research states that hot storage batch read latency dropped by about 5 times compared with DynamoDB's full distribution (P50 from 31.4ms to 5.60ms).
- Perplexity to integrate GPT-6 Astra (2026-09-03): Perplexity announced it will integrate OpenAI's GPT-6 Astra; its CEO cited Astra ranking first on the WANDR benchmark, bringing stronger reasoning and agentic task execution to Perplexity users.
- Perplexity Computer launches Projects (2026-07-30): Launching Projects on Perplexity Computer. With the release of Projects, we are transforming Computer into a multi-agent collaborative operating system for work, featuring persistent memory, files, and session scope across centers and users. Now available to all users!
- Perplexity open-sources Numbat, an agent detection layer (2026-07-29): Today we open-sourced Numbat, an agent detection and response layer designed to work across multiple agent frameworks. Numbat provides security teams with visibility into agent activities and can block selected actions before execution. Learn more: https://research.perplexity.ai/articles/securing-agents-across-perplexity%E2%80%99s-client-endpoints-with-numba
- Perplexity launches a personal computer agent for Windows (2026-07-28): Personal Computer is now available in the Perplexity Windows app. Personal Computer is a local intelligent agent tool designed for your work. It coordinates agents across local files, connected applications, and networks. Research, code, browse, and build—all within a unified system.
- Perplexity + self-developed Sonar / Pages (2025): Self-developed Sonar model + multiple model options (Claude 3.5 Sonnet / GPT-4o / Gemini 2.0), Pages one-click generation of topic pages, over 200,000 enterprise users
- Perplexity 1.0 (2022-2024): Started from "connected AI search," establishing a leading position in the AI search track