Perplexity

Perplexity is an AI-native search company that positions large language models as the front end for information retrieval, in place of the ten-blue-links model that defined Google search. Founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski, it operates an "answer engine" and the Comet agentic browser, and is a defendant in an ongoing news-publisher dispute over scraping, summarization, and attribution.

Snapshot

Valuation

Overview

Rather than training its own frontier models, Perplexity primarily orchestrates over third-party models (OpenAI, Anthropic, Meta, Qwen) and layers proprietary retrieval, ranking, and citation on top. The company's bet is that the interface and retrieval layer is more defensible than the base model, which makes it a test case for whether AI-native search is a durable product category or a feature that hyperscalers absorb.

Products

Comet browser

Comet is Perplexity's entry into the agentic browser category, a product surface in which the browser itself runs persistent AI agents capable of taking actions such as booking, purchasing, and form-filling across web surfaces. It competes with Anthropic's Claude Cowork and desktop agents, OpenAI's Operator and agentic ChatGPT surfaces, and Google's agentic Chrome and Gemini integrations. Agentic browsers raise safety and security concerns, including prompt-injection vectors against user browsing sessions, discussed in Agentic AI.

On June 20, 2026, Perplexity launched Brain, described as a self-improving memory layer for its Computer agent. According to Perplexity, Brain reviews each prior session overnight to identify what succeeded and failed, then begins subsequent tasks faster and at lower cost (Source: perplexity.ai). The release situates Perplexity's agent within a wider set of June 2026 agent-infrastructure launches across the coding-agent and agentic-tooling space.

On July 15, 2026, Perplexity introduced SPACE (Sandboxed Platform for Agentic Code Execution), the sandbox platform behind its Computer agent. According to Perplexity, SPACE runs each task in a Firecracker microVM, keeps user credentials outside the sandbox entirely — passwords and keys stay in user-controlled credential systems — and supports pausing, resuming, and forking agent sessions via rolling snapshots. The company said 100% of Computer agent sessions now run on SPACE, and that the runtime cut median sandbox-creation time from 185ms to 60ms and P90 from 447ms to 89ms (Source: perplexity.ai; research.perplexity.ai). The credential-isolation design addresses the prompt-injection and session-hijack concerns raised about agentic browsing in Agentic AI.

Skills

In May 2026, AI Disruption reported that Perplexity opened up its Skills design, extending the agent-skill pattern adopted by Anthropic (Claude Skills), OpenAI (Codex Skills), and Google (COSMO 14 on-device skills) into Perplexity's answer-engine and Comet stack (Source: aidisruption.ai). The move toward discrete, composable skills in place of single-model dependence aligns with the "12-to-100-model stitching" framing attributed to Marc Andreessen.

News-publisher disputes

Because Perplexity's product summarizes web content as its core value proposition, the company is more exposed on the news-scraping axis than chat-first products such as ChatGPT or Claude. The disputes form part of the broader AI copyright litigation landscape, and the underlying question of whether AI summarization of news content is fair use is the narrower version of the question raised in NYT v. OpenAI.

In 2024, Forbes, Wired, and Condé Nast lodged publisher complaints alleging that Perplexity reproduced content verbatim, mis-attributed quotes, or generated summaries with inadequate citation; Wired's parent Condé Nast issued a cease-and-desist. The New York Times sent a cease-and-desist letter in 2024 demanding that Perplexity stop scraping and using NYT content. In October 2024, Dow Jones and News Corp filed a federal suit alleging copyright infringement and "hallucinated" misattribution of content to their publications. Separately, reporting alleged that Perplexity's crawlers ignored or disguised themselves to bypass robots.txt exclusions, a technical practice distinct from the copyright theory of the cases.

Perplexity responded with a publisher revenue-share program, announced in mid-2024, with partners including Time, Der Spiegel, and Fortune, and with stronger citation policies. Whether AI summarization of news content is fair use remains unresolved.

Policy positions

Perplexity has argued publicly that retrieval-augmented summarization with citation is squarely within fair use, a position the publisher plaintiffs dispute. Like most application-layer companies, it has not issued a responsible-scaling policy or frontier-safety framework, relying instead on the underlying model providers, and it has not been a prominent voice in pause or slowdown discourse.

People

• Aravind Srinivas (CEO, co-founder): Former OpenAI and DeepMind researcher; UC Berkeley PhD.
• Denis Yarats (co-founder, CTO): Ex-Meta AI researcher.
• Johnny Ho (co-founder): Engineering.
• Andy Konwinski (co-founder): Previously co-founded Databricks.

Controversies

The Pages feature, which generates Wikipedia-style reference articles, drew sharp publisher criticism because the generated content competed directly with the source publications. Perplexity has also produced fabricated citations in multiple instances, citing sources that did not support the claim or fabricating attribution, an instance of the hallucination concern from Sycophancy and Hallucination arising in a product where citation is the core value proposition. Wired reporting in June 2024 alleged that the Perplexity crawler disguised itself to bypass robots.txt exclusions.

Relationships

• instance-of: Agentic AI (Comet); AI Diffusion (application-layer)
• depends-on: OpenAI, Anthropic, Meta AI, Alibaba / Qwen Team (model substrates)
• related: AI Copyright Litigation — Analysis, NYT v. OpenAI and Microsoft — Complaint (Dec 2023), Sycophancy and Hallucination, AI and Content Moderation, AI Is Really Weird
• contradicts: the "publishers must license training data" theory advanced by news-publisher plaintiffs; Perplexity's fair-use position sits at the far end of that debate