Perplexity Computer: What if the Future of AI Isn't Picking One Model?

The Virtuoso Race Just Hit a Fork in the Road

While the AI industry's been locked in a sprint to build the single best model, Perplexity just placed a different bet entirely. Instead of picking a winner, they're conducting the whole orchestra.

Computer, Perplexity's latest offering, doesn't care which model is "best." It orchestrates 19 of them—Claude Opus 4.6, Gemini, GPT-5.2, Grok, and 15 others—routing tasks to whichever model handles that specific job best. Think of it as the conductor who knows when to bring in the strings versus the brass.

The pitch is compelling. Text Aravind Srinivas, Perplexity's CEO, something like "plan my two-week Japan trip," and Computer spins up a background workflow that runs for hours, days, or months. It'll search flights, compare hotels, build itineraries, book reservations—then ping you only when it needs a credit card number or a human decision.

As Srinivas told Fortune: "Even your mom can text and delegate."

Why This Matters Now

The AI model landscape fragmented hard over the past year. In January 2025, two models controlled 90% of the market. By December 2025, no single model held more than 25%. What looked like a race to convergence turned out to be a race to specialization.

Claude's great at reasoning and code. GPT excels at general knowledge. Gemini crushes multimodal tasks. Grok... well, Grok does Grok things. Rather than one model learning to do everything, we're watching models carve out niches.

Computer's thesis: stop fighting that trend. Embrace it.

The Three Approaches (According to The Verge)

The Verge framed three competing visions for how AI agents should work:

OpenClaw: Runs locally on your machine. Full control, zero cloud dependency, maximum power. Also maximum risk—it's got the keys to your filesystem and can execute anything. One mistake, one hallucination, and you're restoring from backup.

Claude Cowork: Cloud-based, sandboxed, safe. But locked to one model (Claude). If your task needs something Claude isn't great at—say, image generation or real-time web scraping—you're out of luck.

Computer: Cloud isolation meets model pluralism. It runs in Perplexity's secure environment, so it can't trash your machine. But it taps 19 models, so if Claude's not the right tool, Computer just routes to Gemini or GPT-5.2 instead.

The tradeoff: you're trusting Perplexity with your workflows, and you're paying $200/month for the privilege.

The Limits Nobody's Talking About Yet

Computer's compelling on paper, but the cracks are visible if you squint.

Price: $200/month isn't enterprise SaaS pricing, it's "exclude casual users" pricing. Currently limited to Perplexity Max subscribers, with Pro and Enterprise access coming later. That's a steep gate for a product that's still proving itself.

Dependency hell: Orchestrating 19 models means relying on 19 suppliers. If OpenAI has an outage, GPT tasks stall. If Anthropic changes pricing, margins compress. If Google deprecates an API, workflows break. Computer's resilience is only as strong as its weakest vendor.

Supervision paradox: The whole pitch is "set it and forget it." But how do you audit a workflow that runs for three days across 19 models? When something goes wrong—wrong hotel, missed flight, budget overrun—tracing the failure across that many handoffs is a nightmare.

Legal exposure: Perplexity's already in court with Dow Jones, the New York Times, and Amazon over copyright and content scraping. Computer escalates that risk. If it's autonomously searching, synthesizing, and acting on copyrighted material for months at a time, the legal surface area just exploded.

The Open Question

Is the future of AI one perfect model, or an ecosystem of specialists?

Computer's betting on the latter. The market's moving that way too—model fragmentation, not convergence. But orchestration introduces complexity, cost, and dependency that single-model approaches avoid.

OpenClaw says: own your stack, run it locally, accept the risk. Claude Cowork says: pick the best single model, sandbox it, keep it simple. Computer says: why choose when you can coordinate all of them?

The answer probably isn't binary. But for now, Perplexity's offering the only product that treats "which model?" as the wrong question.

Whether that's prescient or premature depends on whether specialization wins—or whether one model eventually eats the whole stack.