Perplexity turns AI search into an agent that builds Excel files and checks the data itself

Perplexity Computer showed how an AI tool can move from being a "smart search" to acting as an executor. In a test on a real research task, the system built a four-sheet Excel file in 7 minutes 7 seconds, cited 33 sources and separately flagged ten benchmark discrepancies.

How the test went

The user asked the system to collect all AI models released between March 7 and March 14, 2026, find benchmark results for each from at least three independent sources, flag figures that did not match, and then calculate the price-to-performance ratio. The output was not just a list of links but a working file with a ranking of seven models, a list of disputed metrics, and explanations of why different publications report different values.

The most interesting part was not the Excel file itself but how it was built. Computer did not copy text from the browser into cells by hand, and it did not squeeze the result into one long answer. It wrote a Python script, build_spreadsheet.py, ran it, hit errors, fixed them itself and rebuilt the file. In other words, the system did not stop at searching and summarizing: it created an artifact that you can open, check and use in your work right away.

Why this is not search

According to Perplexity, Computer is an agent layer launched on February 25, 2026 on top of a set of 19 models. Claude Opus 4.6 handles central orchestration and coding tasks, deep research goes to Gemini, quick fact checks go to Grok, and specific media tasks are routed to other models. The user does not give a chain of commands but a goal, after which the system splits it into subtasks, distributes the work and assembles the final result.

"Perplexity is no longer a search engine.

It is a different category of tool."

The test showed this in the details. For example, Computer found a gap in GPT-5.4's GPQA Diamond score: 92.8% in OpenAI's blog versus 74.8% at NxCode, and tied the discrepancy to different reasoning effort settings. Even more useful was its cost finding: NVIDIA's Nemotron 3 Super, priced at $0.10/$0.50 per million tokens, came out at a cost/performance figure of $0.37 versus $9.17 for GPT-5.4. The system made this calculation on its own, without a separate request.

What was announced next

On March 11, 2026, at the Ask 2026 conference, the company showed that Computer is not a one-off demo feature but part of a broader product. In the personal version, the agent can run on a dedicated Mac mini, keep permanent access to local files and connected services, watch for triggers and run tasks in the background. Sensitive actions require user confirmation, and a kill switch gives quick manual control.

The company also showed how it plans to monetize and extend this approach. Access to Personal Computer is open through a waitlist for the Max plan at $200 per month, while the enterprise version can already connect to Snowflake, Databricks, Salesforce and HubSpot. For developers, the company exposed the same internal components that power Computer, and it added paid databases and mobile sync to the research setup. The key points of the announcement:

• Personal Computer with persistent access to files, apps and sessions
• Enterprise version with connectors to Snowflake, Databricks, Salesforce and HubSpot
• Four developer APIs: Search, Agent, Embeddings and Sandbox
• Direct access to Statista, CB Insights and PitchBook data inside research
• iOS launch with cross-device sync, followed by Android

What this means

The main shift is that the value of AI is measured less and less by a polished chat answer and more by a finished work product: a table, a report, a calculation, a list of disputed data. Perplexity still has weak spots: an opaque credit system, dependence on the Apple ecosystem for the personal version, and unverified enterprise figures. But the trajectory is clear: AI search is turning into a research agent that does not just find, but also does.