Perplexity Bets on Model Orchestration

This is a bet that the winning prosumer agent will be the one that picks the best model for each subtask, not the one that forces every task through one lab stack. Perplexity had already seen AI search with citations get copied by ChatGPT, Claude, Gemini, and Copilot, so it moved from answering questions to running long workflows across research, dashboards, content, and app actions, then monetized that shift with Computer Credits layered on top of high end subscriptions.1 2 3 4 5

• Computer is not just another chat surface. It was launched in February 2026 as a cloud native agent that can stay alive for long running work, use multiple specialist models, and connect to outside tools. That architecture matches knowledge work where research, drafting, images, video, and app actions happen in the same job.1 6 7
• The competitive logic is different from OpenAI and Google. OpenAI has been pushing vertical integration to own distribution, usage, and training data, including its reported bid for Windsurf. Perplexity is taking the opposite side, arguing that orchestration quality and workflow data can matter more than owning the base model when tasks span many tools.3 4 8
• The hard part is reliability, not raw model IQ. In real Cowork usage, one to three connected tools can work well, but four or five tool chains often break because bad data, missing permissions, or one hallucinated step cascades through Slack, Airtable, project tools, and internal systems. That makes debugging, audit trails, and human review core product features.5 9

Where this heads next is a split between labs that own the whole stack and product companies that win by routing across stacks better than the labs do themselves. If Perplexity keeps turning orchestration into a smoother, less technical workflow, its moat becomes accumulated user context, tool connections, and workflow traces that make Computer harder to swap out than a standalone model.2 5 9