Perplexity has fundamentally changed how AI systems tackle hard research questions by distributing the work across multiple specialized models instead of relying on a single AI engine. The company moved its Deep Research feature into Computer, a cloud-based orchestration system that breaks down complex questions into smaller subtasks and routes each one to the model best suited for the job. The results are striking: on a benchmark measuring an AI's ability to find hard-to-locate information through web browsing, accuracy jumped from 40.7% to 83.8%.

What Changed in Perplexity's Deep Research System?

Deep Research is a mode that runs multiple searches, reads sources, and writes a cited report. The new version, launched inside Perplexity Computer in late February 2026, represents a shift from single-model reasoning to what Perplexity calls "Search as Code." Instead of following a fixed pipeline of steps, the system now writes code that automatically designs its own search strategy tailored to each question.

The core innovation is parallel execution. The code-driven search runs thousands of retrieval steps simultaneously, each one adapted to the specific question being asked. This differs fundamentally from traditional search, which applies the same sequence of steps to every query. The system uses Perplexity's Agentic Search SDK, which exposes search primitives like filtering, deduplication, and reranking that the AI can combine in novel ways.

Computer itself coordinates up to 20 AI models in a single workflow, with Anthropic's Claude Opus 4.6 serving as the core reasoning engine. For specialized tasks, the system routes work to other models; for example, Google's Gemini handles deep research tasks while other models focus on legal reasoning, data analysis, or writing.

How Does Multi-Model Routing Improve Research Quality?

The benchmark gains reveal where multi-model orchestration delivers the biggest wins. On BrowseComp, a test that measures an AI's ability to navigate multiple web pages to find obscure information, the jump from 40.7% to 83.8% is the largest improvement shown. On Humanity's Last Exam, which covers expert-level questions across academic subjects, accuracy rose from 36.4% to 50.5%. DeepSearchQA, which already performed well, saw a smaller but still positive gain.

The system reads your internal files alongside the live web, cross-referencing a PDF or spreadsheet against census data, Statista, and other sources. Every claim in the final report is cited inline, creating a transparent chain of evidence. The deliverables are work-ready: reports, briefs, decks, dashboards, and live spreadsheets that Computer can read and write directly.

How to Use Deep Research in Computer for Real-World Tasks

• Finance Analysis: Compare cash flow and profit margins of major AI chip companies over five years, with Computer pulling data from premium sources like PitchBook and CB Insights.
• Legal Compliance: Map how US and European data-privacy laws differ into a single comparison table, with legal reasoning models handling contract review and regulatory interpretation.
• Healthcare Research: Synthesize clinical-trial evidence on whether weight-loss drugs improve heart health, aggregating findings from multiple studies with cited sources.
• Technology Benchmarking: Evaluate leading AI models on reasoning ability, cost, and context length, with data models handling spreadsheet variance checks and comparisons.

Each task ends in a deliverable that you can approve or reject before it lands. Computer shows a preview of any changes it plans to make, giving users control over the final output.

Who Can Access This Feature and How Much Does It Cost?

Deep Research in Computer is a consumer feature available to Perplexity Max subscribers. Developers can access the same agentic search stack through Perplexity's pay-as-you-go Agent API, which ships a deep-research preset. The API endpoint is POST https://api.perplexity.ai/v1/agent and also accepts POST /v1/responses for OpenAI SDK compatibility.

The system represents a broader shift in how AI systems handle complex reasoning. Rather than asking a single model to do everything, the industry is moving toward orchestration, where different models specialize in different tasks and a coordinator decides who does what. This approach mirrors how human teams work: a lawyer handles contracts, a data analyst handles spreadsheets, and a writer handles the final draft.

What Are the Limitations?

Perplexity published the benchmark numbers itself, so independent verification still matters. Premium-source coverage varies by domain, and legal data remains in preview. Most importantly, outputs still require human review; being cited does not always mean correct. The system is designed to augment human researchers, not replace them.

The broader context matters too. A separate survey on reasoning and agentic systems in time series data found that LLMs (large language models) are increasingly being used to execute structured reasoning procedures over temporally indexed data, enriched by multimodal context and agentic systems. This reflects a wider industry trend toward systems that not only analyze but also explain, reason about counterfactuals, and decide among alternative actions.

Perplexity's move signals that the future of AI research tools lies not in building a single all-powerful model, but in orchestrating multiple specialized models to work in parallel. The 83.8% accuracy on information-finding tasks suggests this approach is working, at least for the kinds of research questions that require navigating multiple sources and synthesizing evidence.