Perplexity’s return and the rise of Model Orchestration

Perplexity just announced Perplexity Computer, a general purpose digital worker that doesn't only answer questions, but also orchestrates entire workflows. Perplexity used to position itself basically as an AI-powered search engine, but now it looks like they are betting to be a 'full-stack' autonomous system capable of researching, designing, coding, deploying, and even managing projects from start to finish. Their philosophy is, today's frontier models are already powerful enough to do extraordinary work, but chat interfaces only expose a fraction of that potential.

The most (technically at least) interesting aspect of the product is its multi-model orchestration engine. Rather than betting on a single model family, Perplexity Computer routes tasks across up to 19 different models, using Claude Opus as its core reasoning layer to match each subtask to whichever specialist performs best. Need video? Veo handles it. Lightweight speed tasks? Grok. Long-context recall? ChatGPT. Research? Gemini. The idea is that the system spawns sub-agents that run in isolated compute environments with access to real browsers, files, and connected tools, and if a subtask stalls, it just spawns more agents to troubleshoot on the fly. Work runs asynchronously, which means users can step away while the system keeps going, a very popular concept nowadays, for better or worse.

This, naturally, invited users to compare with the viral personal AI agent OpenClaw, an assistant you run on your own machine and actually executes tasks you ask it to do. And yeah, on the surface they share some similarities (both can perform multi-step tasks and aim to go beyond simple Q&A, for example) but their philosophies differ a lot. OpenClaw is focused on giving a single user persistent agency over their own environment via chat, whereas Perplexity Computer is explicitly trying to be a project-wide autonomous orchestration platform. It isn’t just an “AI assistant inside your browser”, since it’s designed to be a memory-aware system that can coordinate multiple specialist models across an entire workflow. Their approaches to model diversity also contrast: Perplexity leans into mix-matching models to suit each subtask, while OpenClaw runs as a user-controlled agent that may integrate with a single model backend.

Online reactions are already reflecting a mix of genuine excitement and, of course, skepticism. Some have noted that recent changes to Perplexity's product - consolidations that felt like cutbacks at the time - now make strategic sense for this pivot. Others who already use Perplexity for research found it solid for quick summaries but expressed hesitation about connecting it to "everything". The guardrails question has also surfaced prominently: accessibility is great and all, but at this level of autonomous action (agents writing code, making API calls, managing files, etc.) safety boundaries matter. Not to mention that for now, Perplexity Computer is available only to Max users, with a pricing point that is not exactly very accesible (although it's coming soon to Pro users).

At a deeper level, Perplexity Computer tells something broader about where the industry is heading and that we’ve been noticing for a while now: the real competitive moat may no longer (only) be which company has the best single model, but which platform can coordinate the best combination of models for any given job. In practice, this reflects a shift from thinking of AI products as isolated assistants toward platforms that integrate models and workflows into strong systems that can act autonomously across long-running tasks. Whether Perplexity can actually pull this off operationally - managing 19 models, persistent memory, hundreds of connectors, and most importantly, real-world reliability - is the actual test that is yet to be seen.