Perplexity creates an environment for AI-driven processes

On May 29, 2025, Perplexity launched Labs , an advanced tool in the Pro subscription that turns text queries into ready-to-use reports, tables, visualizations, and even mini web apps. Labs doesn’t just interpret commands: it performs web research, generates and executes Python or JavaScript code, visualizes data using libraries like Plotly and D3.js, and then exports the results to DOCX, CSV, HTML, or as embeddable widgets.

The tool works in a browser and is also available on iOS and Android, with plans to launch on macOS and Windows by the end of the year. With one command in the chat, the user can, for example, build a regression model based on open data, visualize sales trends by region, or put together a landing page for an A/B test.

On July 9, the company unveiled Comet , a standalone Chromium-based browser with an AI assistant built in that maintains global context between tabs and executes action-level commands, from “place an order” to “book an appointment,” without user intervention.

Comet is currently only available to Max subscribers and a limited number of users on a waitlist, with a general free release promised later.

Comet already supports integrations with Gmail, Google Calendar, LinkedIn and is able to search and filter candidate profiles, send personalized emails, and automatically book calendar slots.

Perplexity CEO Aravind Srinivas describes Comet as an attempt to transform the browser into an “AI operating system” that can replace entire roles—recruiters, administrative assistants, analysts—without losing accuracy or productivity.

Under the hood, Perplexity uses a hybrid architecture: a combination of models from OpenAI (GPT-4o), Anthropic (Claude 3), and its own Sonar line. Models are selected dynamically - depending on the request, the type of task, and the optimal cost of execution.

This modular approach provides up to 28% increase in accuracy on complex analytical queries (according to internal A/B tests), a 12–18% reduction in average query cost due to adaptive routing between LLMs, and stable benchmark coverage:

• MMLU : 86.2 (Sonar-Large) - an indicator of comparative analysis and identification of weak points of LLM in real subject problems.
• HumanEval (Python codegen) : 76.1% - shows how "smart" the language model is when solving practical programming problems in Python.
• BrowserQA (personalized tasks) : 82.4 - shows whether the AI agent can work in the browser just like a human: follow links, extract data and give an accurate result.

According to Kirill Pshinnik, co-founder of Zerocoder University, "the Labs and Comet tools remove the barrier between intent and implementation. Tasks that previously required a combination of "human + tools" (for example, running a script, collecting data, analyzing it, and visualizing it) are now solved with one phrase in the browser."

This opens up a completely new workflow:

• the marketer launches a hypothesis without contacting the developer,
• HR generates analytics on the job market,
• The designer publishes a landing page with automatic code generation and SEO.

But as AI takes on the role of email and calendar operator, privacy issues arise. Automating routine roles could also lead to a reshaping of office functions. Still, the direction is clear: the browser is becoming a digital employee, and analytics is becoming a self-service tool.

But Kirill Pshinnik adds that “for IT specialists, marketers, product managers and creative teams, this is a signal: whoever quickly implements an AI agent into their work will quickly free up resources for tasks in which humans are still irreplaceable – strategy, creativity, decision-making.”