Perplexity’s Deep Research Upgrade Now Routes Subtasks Across 20+ AI Models
Perplexity has embedded its Deep Research tool inside Computer, its multi-model orchestration layer, transforming how complex queries are handled. The update splits tough questions into specialized subtasks, distributes them across more than 20 frontier models, and delivers ready-to-use reports, slide decks, and dashboards—all within the same workflow.
A smarter way to tackle complex research
Deep Research now runs inside Perplexity Computer, which debuted in late February 2026. Computer acts as a cloud orchestrator that can coordinate up to 20 AI models in a single workflow, independent of any specific provider. Opus 4.6 serves as the core reasoning engine, while specialized sub-agents—such as Gemini for deep research—handle targeted tasks. The system automatically drafts a research plan, gathers primary sources from hundreds of sites, and cites every claim, ensuring transparency and traceability.
Code-driven searches that adapt on the fly
A key innovation is “Search as Code,” which lets the model write code that dynamically builds each search. Instead of following a fixed pipeline, the script runs thousands of retrieval steps in parallel, tailored to the question at hand. It runs in a sandboxed environment and taps into Perplexity’s Agentic Search SDK, which exposes primitives like filtering, deduplication, and reranking. This approach allows the system to branch, compare, and refine its approach as new information emerges. The same stack is also accessible to developers through the pay-as-you-go Agent API, enabling programmatic access to the agentic search capabilities.
From benchmarks to real-world use
Perplexity reports noticeable gains over its legacy Deep Research. On BrowseComp, a benchmark testing an agent’s ability to locate hard-to-find information, performance jumped from 40.7% to 83.8%. Smaller but still positive improvements appear on other tests such as Humanity’s Last Exam and DeepSearchQA. Computer can also ingest local files—PDFs, spreadsheets—alongside live web sources, cross-referencing internal data against public datasets like census records and Statista. A developer example shows how to spin up a deep-research session via the Perplexity SDK, comparing cash flow and profit margins of major AI chip makers.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.