June 2026 Summaries

Perplexity has launched Computer for Counsel, an AI product designed to reduce lawyers’ administrative workload by connecting Perplexity Computer with legal research databases, document repositories, contract tools, and matter-management systems. Powered by more than 20 AI models, it selects models for tasks such as legal research, document collection, contract review, drafting, and regulatory monitoring while providing linked, verifiable citations from sources including Midpage, Deel, and, soon, LegalZoom. The product integrates with Microsoft 365, Google Workspace, and hundreds of connected tools, including Box, Docusign, NetDocuments, DeepJudge, and others, enabling it to assemble relevant internal and external information into briefs, memos, deal summaries, and workflow outputs. Perplexity says attorneys retain responsibility for legal judgment and strategy, while enterprise data is not used for model training and connected files remain under customer control. Computer for Counsel is available to Perplexity Enterprise and Max subscribers.

Brain is a new self-improving memory system for Perplexity’s Computer agent that focuses on remembering and learning from work performed rather than primarily retaining user preferences. It creates a traceable context graph, presented as an automatically updated LLM wiki, that incorporates past sessions, connected sources, documents, project details, successful approaches, mistakes, and user corrections. By reviewing and synthesizing this information at intervals such as overnight, Brain aims to help agents begin future tasks with stronger context, identify reliable sources more quickly, reduce unnecessary model calls and token use, and improve outputs over time. Early measurements cited by the company indicate that Brain increases correctness by 25% and recall by 16% on previously encountered tasks while reducing costs for historically contextual tasks by 13%. Memory entries link to their originating sessions, files, or sources for traceability, and Brain is initially available in Research Preview to Max and Enterprise Max subscribers.

Deep Research has been integrated into Computer to provide an agentic research workflow that plans queries, performs iterative and parallel searches, evaluates evidence and conflicts, incorporates internal and external sources, and converts findings into deliverables such as reports, dashboards, PDFs, and presentations. The new “Search as Code” architecture enables the system to write and execute custom retrieval programs in a secure sandbox, potentially running thousands of searches, filtering and deduplicating results, and adapting its strategy as evidence emerges. Computer can route complex questions across more than 20 models and combine legal, financial, scientific, or business context into a unified, cited response, including premium sources such as Statista, PitchBook, and CB Insights when appropriate. The announcement cites benchmark improvements in factual accuracy, analytical depth, and citation quality, while highlighting applications including investment due diligence, healthcare competitive analysis, cybersecurity audits, and multijurisdictional legal assessments.

A study conducted with Harvard Business School researchers examines how users employ Perplexity Computer, an agent-based system that coordinates more than 20 AI models and connected tools to complete multi-step knowledge-work tasks from high-level objectives. Compared with Perplexity Search on similar requests, Computer performed substantially more machine work and external-tool interactions, while the study estimates it reduced task completion time by 87% and cost by 94% across 10,000 matched tasks spanning 18 fields. Users increasingly requested finished outputs rather than answers, with many tasks involving creation, analysis, non-routine work, and multiple areas of expertise, including work outside users’ primary occupations. The findings suggest that by automating execution, users shift toward defining goals and reviewing results, enabling them to pursue work that may previously have required more time or specialized support; however, the authors note that Computer is still new and that its longer-term effects on teams and organizations remain uncertain.

A Perplexity and Harvard Business School study examines how its Computer agent, launched in February 2026, differs from conventional AI search by autonomously planning and executing multistep work across tools, while users focus more on defining goals and reviewing results. Based on product-usage data, matched Search and Computer sessions, task classifications, and user interviews, the study reports that Computer use grew rapidly and spans research, analysis, document creation, software, finance, operations, and other domains. In matched tasks, Computer averaged 26 minutes of machine execution compared with 33 seconds for Search, used more external connectors, and showed lower measured user dissatisfaction despite occasionally pausing for clarification or approval. The authors estimate that Computer can reduce total task time by 84–87% and estimated costs by 93–94%, though these figures depend on assumptions about human labor, oversight, and equivalent execution time. Computer users also more often attempt work outside their inferred occupations, pursue more complex and non-routine tasks, and engage a broader range of knowledge areas and work activities, particularly in software development, document production, and visualization. The study cautions that its early-adopter sample, session-matching approach, estimation methods, and focus on the Perplexity ecosystem limit generalizability, while suggesting that agents may reshape knowledge work by shifting people from manual operators to supervisors and potentially changing how roles and teams are organized.

Perplexity has announced the Main Street AI Accelerator, a $25 million Computer credits program coordinated with Small Business Development Centers of the U.S. Small Business Administration, aimed at helping up to 100,000 eligible growing businesses access $250 in credits. The initiative reflects the view that startups and traditional small businesses increasingly rely on similar digital tools and face common challenges in managing fragmented software and underused information. Perplexity’s Computer platform connects with more than 400 business applications, including QuickBooks, Mailchimp, Shopify, Stripe, Supabase, Vercel, Apollo, Canva, and Semrush, spanning operations such as business formation, product development, sales, marketing, and customer management. For specialized tools without direct integrations, Perplexity says its Comet agentic browser can log into cloud-based services, retrieve data, and complete tasks, positioning Computer and Comet together as an operational layer for businesses’ existing software.

Perplexity has announced Personal Computer for Windows, extending its AI agent system to let users work across local files, Microsoft applications, connected business tools, and the web from a Windows device. Following launches for Microsoft 365 and Mac, the product uses multi-model orchestration to carry out tasks such as updating Word, Excel, and PowerPoint files, organizing downloads, researching current information, handling emails, and completing web-based actions, with support for more than 400 tools including Snowflake, Salesforce, and HubSpot. It also supports cross-device workflows, allowing users to begin tasks on a phone and have their Windows PC complete them, and offers voice controls for direct commands. The company highlights use cases in finance, legal, and sales, while stating that sensitive actions require user alerts and that files and activities are handled through a secure, auditable, and reversible sandbox environment. Personal Computer for Windows will initially be available through a waitlist for paying Max and Enterprise Max subscribers.

Perplexity announces Personal Computer, described as a hybrid local-server inference orchestrator that automatically divides AI tasks between a user’s device and cloud-based frontier models to balance accuracy, privacy, cost, and energy use. A compact local model identifies sensitive information, such as financial, health, or personal data, that should remain on-device, while more demanding components can be processed in the cloud. The company says its model-agnostic system builds on its existing orchestration of tools, sources, and AI agents, and will support local hardware including Intel chips and NVIDIA RTX Spark. Perplexity argues that stronger local compute could reduce reliance on centralized data centers, support data sovereignty, and improve efficiency by reserving cloud resources for tasks that require them. Personal Computer with local inference is planned for release in July.

Perplexity introduces Search as Code (SaC), an agent-oriented search architecture that replaces fixed, monolithic search pipelines with modular retrieval, ranking, filtering, fan-out, and rendering primitives exposed through a Python-based Agentic Search SDK. Rather than issuing serial function calls and receiving preprocessed result sets, models generate and execute task-specific code in secure sandboxes, enabling asynchronous workflows, thousands of retrieval operations, deterministic processing, and more selective use of context. SaC combines models as planning and control layers, sandboxes for code execution and state management, and an SDK built from atomized search components; Perplexity reports favoring filesystem-based state serialization over persistent REPL environments for long-running tasks. A CVE-advisory case study illustrates how agents can generate targeted vendor searches, adapt coverage strategies, and verify evidence linking vulnerabilities to product fixes. Across five benchmarks, including DSQA, BrowseComp, HLE, WideSearch, and the new WANDR benchmark, Perplexity reports that its SaC implementation outperformed competing evaluated systems on four benchmarks and was effectively tied on HLE, while also showing favorable cost-performance results. The company frames SaC as a hybrid computing model in which language-model reasoning determines research strategies while deterministic code handles parallelism, filtering, aggregation, and other operational tasks.