Perplexity Computer has been announced by the AI company as a "general-purpose digital worker."

Perplexity has introduced Perplexity Computer, a new system designed to move beyond traditional AI chatbots and task-based agents.

Instead of just answering questions or completing one-off actions, Perplexity Computer is built to handle entire workflows from start to finish.

With Perplexity Computer, you describe the outcome you want, and the system figures out how to get there. It breaks the goal into smaller tasks and subtasks, assigns them to specialised sub-agents, and executes them in the background. One agent might gather information from the web, another could draft documents, process data, or connect to external services through APIs. The coordination happens automatically, and the work runs asynchronously, meaning users don't have to stay actively involved throughout.

Perplexity Computer operates much like a human colleague would, by using software tools directly. It works within a real browser, accesses files, writes code when needed, and interacts with connected services. If it runs into a roadblock, it can create additional sub-agents to solve the issue, whether that involves researching missing information, locating credentials, or asking the user for clarification when necessary. Each task runs inside an isolated compute environment for controlled execution.

A major part of the system is its multi-model approach. Rather than relying on a single AI model, Perplexity Computer orchestrates different models depending on the job. Some models are used for core reasoning, others for research, image generation, video tasks, faster lightweight operations, or long-context recall.

Also Read: OpenAI Starts Testing Ads in ChatGPT

As of now, Perplexity Computer runs Opus 4.6 for its core reasoning engine and orchestrates sub-agents with the best models for specific tasks: Gemini for deep research (creating sub-agents), Nano Banana for images, Veo 3.1 for video, Grok for speed in lightweight tasks, and ChatGPT 5.2 for long-context recall and wide search.

The framework is designed to remain flexible, allowing models to be swapped out as newer ones emerge. Users can also choose specific models for specific subtasks, which becomes important as compute costs and token limits factor into real-world usage.