Perplexity Computer puts 19 models to work — if it works

Perplexity shipped a new product in late February called Perplexity Computer. It’s not a physical device. It’s a cloud-based agent that orchestrates 19 different AI models to complete complex workflows on your behalf — booking travel, researching competitors, managing multi-step tasks that span a dozen apps.

Hours before the press briefing to announce it, Perplexity canceled the demo. They’d found flaws in the product.

The pitch and the problem

The premise is ambitious: rather than betting on a single frontier model, Perplexity Computer creates “subagents” that spin up specialized models for specific sub-problems within a larger workflow. Need visual output? The system might route to Gemini Flash. Software engineering task? Claude Sonnet 4.5. Medical research? GPT-5.1. These were the actual user preferences Perplexity observed in December 2025.

Multi-model is the future.

The product is available only to Perplexity Max subscribers at $200 per month — a tier launched in July 2025 that also includes early access to the Comet browser and priority access to frontier models. At that price point, Perplexity isn’t chasing volume. They’re explicit about this. “You don’t hear us talk about MAUs ever,” one executive said at the briefing, “because we’re not actually on a mission to get as many users as possible.”

The target user: someone making “GDP-moving decisions.”

The strategic pivot

A company in transition. Perplexity started as an AI-powered search engine, essentially a wrapper around other companies’ models and web indexes. In 2025, they expanded into Labs (spreadsheet and report generation), launched the Comet browser with its own assistant, and built their own AI-optimized search API so they no longer depend on third-party web indexes.

Late in 2025, they abandoned advertising entirely. The stated reason: trust. When your product promises accurate answers, selling ads against those answers creates a tension that users notice.

The financials explain the urgency. In 2024, Perplexity burned approximately $65 million against revenue of $34 million. By January 2025, annualized recurring revenue had reached $80 million, and they were raising $500 million at a $14 billion valuation. But they’re competing against OpenAI’s claimed 800 million weekly users. Perplexity’s total user base is “tens of millions.”

The path to survival runs through enterprise revenue, not consumer growth.

The execution gap

Perplexity Computer represents the company’s bet that orchestration beats scale — that coordinating many specialized models is more valuable than one dominant model doing everything. It’s a reasonable thesis. The December 2025 user data showing different model preferences for different task types supports it.

But the demo cancellation tells a different story. When a company pulls its own product demonstration hours before showing it to press, something fundamental isn’t working. TechCrunch’s earlier hands-on testing with Comet Assistant found a similar pattern: “helpful for simple tasks” but “quickly falls apart with more complex requests” and “hallucinated” during a booking task.

Multi-model orchestration may indeed be the future. But orchestrating 19 models means nineteen points of potential failure, nineteen latencies to manage, nineteen context windows to coordinate. The ambition is evident. The execution, as of late February, was not ready for a live demo.

What to watch

Perplexity’s developer conference, called “Ask,” was scheduled for March 11 in San Francisco. Comet is coming to iOS this month. The company is betting its future on agentic tools for the enterprise rather than consumer search. If Perplexity Computer stabilizes and delivers on its multi-model promise, the $200/month tier becomes a serious wedge into enterprise workflows. If it doesn’t, the demo cancellation becomes the footnote to a familiar story: the gap between what AI agents promise and what they actually do.

The lesson from Humane’s AI Pin — a hardware product that launched “thoroughly unfinished and totally broken” in 2024 — is that users don’t forgive products shipped before they work. Perplexity, at least, had the sense to cancel the demo. Whether they have time to fix what broke is the open question.