Perplexity Moves Deep Research Into Computer, Routing Tasks Across 20+ AI Models

Perplexity has moved its Deep Research capability into Computer — its multi‑model orchestration system — in a quiet upgrade that changes how the tool handles complex research questions. Instead of running Deep Research as a standalone mode, Perplexity now breaks hard questions into subtasks and routes them across 20+ frontier AI models, returning work‑ready reports, decks, and dashboards, according to MarkTechPost.

The new version lives inside Perplexity Computer, which launched in late February 2026. Computer is a cloud system that coordinates up to 20 AI models in one workflow, operating model‑agnostically with Claude Opus 4.6 as its core reasoning engine and specialized sub‑agents handling specific tasks — Gemini for deep research, GPT‑5.2 for long‑context recall, Grok for lightweight speed‑sensitive work, and others for images and video.

Perplexity Computer doesn't just pick one model and run with it. When you ask a complex question, the system decomposes it into tasks and subtasks, creates sub‑agents for execution, and routes each piece to the model best suited for it. One agent might research while another gathers data, a third writes while a fourth generates charts — all coordinated automatically.

According to Till Freitag's analysis, the system runs on 19 models under the hood, with the roster changing as new models prove their strengths. The model assignments are not fixed: Claude Opus 4.6 handles orchestration and coding, Gemini powers deep research, GPT‑5.2 manages long‑context recall, Grok handles lightweight tasks, and Nano Banana and Veo 3.1 handle images and video respectively.

Perplexity is betting on a contrarian thesis: AI models are specializing, not commoditizing. The company shared internal data showing that in January 2025, over 90% of enterprise tasks on its platform were handled by just two models. By December 2025, no single model commanded more than 25% of usage, as reported by Till Freitag.

A Perplexity executive put it bluntly in the same analysis: Claude Opus 4.6 is "a terrible writer" — but it's the best coder available. A marketing team using Claude will underperform one using Gemini. An engineering team using Gemini will underperform one using Claude. No single model serves all use cases equally well, and Perplexity's orchestration layer is designed to exploit those differences.

The Deep Research upgrade is built on two components: the Agent Search SDK and Search as Code. When you submit a complex question, the system automatically builds a research plan, finds primary sources across hundreds of sites, and cites every claim. The output is no longer just a text report — Computer now produces reports, slide decks, and dashboards inside the same workflow.

The key difference from standalone Deep Research is that each subtask now gets its own specialized model. If one part of the research requires deep analytical reading, it goes to a model optimized for that. If another part needs broad web search and synthesis, it gets routed differently. The result, Perplexity claims, is improved accuracy, depth of analysis, and citation quality.

Perplexity's approach stands in contrast to competitors betting on vertical integration. OpenAI and Anthropic are building ecosystems around their own models — ChatGPT and Claude handle everything internally. Perplexity is building the horizontal orchestration layer that sits on top of everyone's models.

² compares it to cloud computing: the companies that built abstraction layers above commodity infrastructure (Kubernetes, Terraform) often captured more value than the infrastructure providers themselves. Perplexity is making the same bet for AI — that the orchestration layer, not the model layer, is where long‑term value accrues.

The unnamed Perplexity executive was quoted in ² of the platform. For developers and researchers who use AI tools daily, the upgrade means Deep Research can now tackle more complex, multi‑part questions with better accuracy — and deliver results in formats (decks, dashboards) that are immediately useful in team settings. The citation quality improvement is especially relevant for anyone using AI‑generated research in professional contexts.

For builders thinking about AI infrastructure, Perplexity's multi‑model bet raises a strategic question: are you building on a single model stack, or designing for a world where models specialize and your system routes intelligently between them? If Perplexity's usage data is right — and the trend from 90% dual‑model concentration to sub‑25% per model in a single year is dramatic — then model‑agnostic architecture isn't a nice‑to‑have. It's table stakes.