Perplexity's AI agent now does deep research — and it's shockingly good

Perplexity is rolling out a major upgrade to its AI agent system, Computer. Starting today, Pro and Max subscribers can tap into Deep Research as a native skill — and the results are frankly ridiculous.

This isn't your average search enhancement. Computer's version of Deep Research is built on a new "Search as Code" architecture and an Agent Search SDK. During a query, the model writes its own code to construct a search pipeline, then fires off thousands of parallel retrieval steps. It's a custom search engine for every single question.

You can feed it local files and apps for internal context, and it'll cross-reference those against live web data in real time. Need to analyze a contract? It'll dispatch a legal reasoning model. Excel discrepancies? A data analysis model handles that. Final report? A writing model takes over. The system orchestrates over 20 frontier models to break down complex problems into sub-tasks, then assembles the results into deliverables like research reports, presentations, and dashboards.

And the benchmarks are nuts. On OpenAI's BrowseComp test, accuracy jumped from 40.7% to 83.8%. On Humanity's Last Exam — a collaboration between CAIS and Scale AI — it went from 36.4% to 50.5%. And on Google DeepMind's DeepSearchQA, it climbed from 81.9% to 85.0%. Perplexity isn't just iterating; it's leapfrogging.