Perplexity has upgraded its Deep Research feature by moving it into Computer, a cloud-based system that coordinates up to 20 different AI models to answer complex questions more accurately and thoroughly. The new version breaks down hard questions into smaller subtasks, routes each one to the model best suited for it, and produces work-ready reports, decks, and dashboards with cited sources. On a benchmark measuring the ability to find hard-to-locate information through web browsing, accuracy jumped from 40.7% to 83.8%.

How Does Perplexity's Multi-Model System Actually Work?

Deep Research in Computer is built on two core technologies: the Agent Search SDK and Search as Code. When you ask a complex question, the system automatically builds a research plan, then writes code that runs thousands of retrieval steps in parallel. This code-driven approach is fundamentally different from traditional search pipelines that follow the same steps every time. Instead, the system can branch, compare, and refine its approach as it learns more about your question.

The Agent Search SDK exposes search primitives like filtering, deduplication, and reranking, giving the system fine-grained control over how it finds and evaluates sources. Computer also reads your internal files alongside the live web, cross-referencing a PDF or spreadsheet against census data, Statista, and other public sources to provide comprehensive context.

What Are the Performance Gains Across Different Tasks?

Perplexity published benchmark results comparing the legacy Deep Research system with the new Computer version. The improvements vary depending on the task type. On BrowseComp, which tests an agent's ability to navigate multiple web pages to find difficult-to-locate information, accuracy jumped from 40.7% to 83.8%. On Humanity's Last Exam, a benchmark covering expert questions across academic subjects, performance rose from 36.4% to 50.5%. The gains are largest on tasks requiring agentic browsing, where the system must intelligently navigate many pages.

The system routes each subtask to the model best suited for it. A legal reasoning model handles contract review, a data model handles spreadsheet variance checks, and a writing model handles the final draft. Premium data sources back the answers, including PitchBook and CB Insights, though legal data remains in preview.

How to Use Deep Research in Computer for Your Own Work

• Consumer Access: Deep Research in Computer is available to Perplexity Max users as a consumer feature, allowing them to generate reports, briefs, decks, and live dashboards directly within the platform.
• Developer Access: Developers can reach the same agentic search stack through the pay-as-you-go Agent API, which ships a deep-research preset and accepts requests via the endpoint POST https://api.perplexity.ai/v1/agent.
• File Integration: Users can pull in PDFs or spreadsheets for internal context, and Computer will cross-reference that data against live web sources, then show a preview before any changes are made to files.
• Deliverable Formats: The system produces ready-to-use outputs including reports, briefs, decks, dashboards, and live spreadsheets, all with inline citations for every claim.

Perplexity ships starter tasks that demonstrate the intended scope. These include comparing cash flow and profit margins of major AI chip companies over five years, mapping how US and European data-privacy laws differ into one comparison table, synthesizing clinical-trial evidence on whether weight-loss drugs improve heart health, and benchmarking leading models on reasoning ability, cost, and context length.

What Are the Practical Limitations?

While the performance gains are significant, several limitations remain. The benchmark numbers are first-party, meaning independent verification still matters. The in-Computer feature centers on Perplexity Max, not a free tier, so access is limited to paid subscribers. Premium-source coverage varies by industry, and legal data remains in preview. Most importantly, outputs still need human review, since being cited does not always mean the information is correct.

Search as Code is rolling out through both Computer and the Agent API, so developers can reach the same agentic search stack programmatically. The official SDK ships a deep-research preset, and the endpoint also accepts POST requests for OpenAI SDK compatibility, making it accessible to developers already familiar with OpenAI's API structure.

The upgrade represents a shift in how AI systems approach complex research tasks. Rather than relying on a single large language model to answer every question, Computer distributes the work across specialized models, each optimized for different types of reasoning and retrieval. This orchestration approach mirrors how human research teams divide labor, with different specialists handling different aspects of a project.