Perplexity on Wednesday, February 25, 2026 unveiled Computer, a multi-agent orchestration system designed to route complex work across a roster of frontier AI models. The pitch is simple: instead of betting everything on one “do-it-all” agent, Computer breaks a goal into subtasks, assigns each to the best-suited model, and runs them in parallel inside an isolated environment.
Computer is available immediately to Perplexity Max subscribers at $200 per month, and it also marks a notable pricing shift: usage-based billing via credits for consumers. Perplexity says broader access for Pro and Enterprise tiers is coming soon, as the company expands beyond its search roots into longer-running, end-to-end workflows that look more like “AI operations” than Q&A.
Nineteen models, one orchestrator
The headline idea behind Computer is coordination. Rather than forcing one model to brute-force a full workflow, Perplexity positions Computer as a manager that can route subtasks to specialists—research, planning, code generation, debugging, design assets, and deployment steps—without a user manually juggling tools.
Reporting around the launch says Computer orchestrates 19 models, with Claude Opus 4.6 serving as a central reasoning engine that decomposes requests and dispatches work outward. Perplexity’s own subscription documentation shows a broad menu of advanced models available to users, reinforcing the “model-agnostic” framing: the orchestrator decides what to use, and users can override those choices when they want direct control.
The safety context: agents are under scrutiny
Computer arrives in a moment when autonomous agents are under more skepticism than hype. Earlier this month, an always-on assistant called OpenClaw went viral — and then triggered backlash after a widely shared incident involving email deletion fears. The episode became a cautionary tale about what happens when an agent misinterprets intent at scale, especially after it’s been trusted with real accounts and real data.
OpenClaw’s creator, Peter Steinberger, also became a headline on his own: OpenAI CEO Sam Altman publicly praised him and said he was joining OpenAI. But the broader takeaway wasn’t celebratory — it was structural. Agents can be helpful in small, controlled contexts, then behave unpredictably when moved into higher-stakes environments.
Sandboxed by design
Perplexity is leaning into that trust gap with a core architectural claim: Computer runs tasks inside an isolated, secure sandbox. The system can use a browser, a filesystem, and integrations, but it does so in a contained environment meant to reduce blast radius when something goes wrong.
That doesn’t solve every problem. A sandbox can prevent catastrophic side effects — the nightmare scenario of an agent wrecking a real inbox or production account — but it can’t guarantee the output is correct. Disagreements between sub-agents, routing mistakes, and failure modes in long-running tasks are the kinds of edge cases that only surface once a system is used at scale.
The pricing shift: “AI work” starts to look like cloud billing
Perplexity’s bigger signal may be financial. Max subscribers pay $200/month and receive a credit allotment, with options like spending caps and usage controls. The reason is obvious: agents can burn through enormous volumes of compute when they’re running parallel tasks for hours, not minutes.
In other words, this isn’t the economics of a chatbot. It’s closer to cloud compute: workload spikes, bursty usage, and a bill that reflects how much work you actually asked the system to do.
What Perplexity is trying to become
Zoom out and Computer looks less like a feature and more like a platform direction. Over the past year, Perplexity has pushed into products adjacent to search — a browser, assistants, vertical tools — but Computer ties those strands together into an “operating layer” that can sit above models and decide which one does what.
That positioning matters because the rest of the industry has strong incentives to route users toward in-house models. Perplexity doesn’t have a foundation model of its own, which makes “honest broker” a plausible identity — at least in theory. Whether that neutrality holds up against API pricing changes, model deprecations, and competitive pressure is the long-term question.
The short-term question is simpler: can a consumer-facing orchestrator make multi-model workflows feel effortless — without turning autonomy into a reliability and cost-control nightmare?