Perplexity AI shipped a product called Computer on February 27, 2026, billing it as an “independent digital worker” that can manage documents, send emails, conduct research, and coordinate tasks across apps without constant human supervision. The system orchestrates work across 19 AI models in parallel and connects to hundreds of services, promising a kind of always-on automation that runs for as long as a user needs it. But the credit-based billing model that fuels Computer introduces friction that complicates the vision of truly uninterrupted, months-long task execution.
What Computer Actually Does
Perplexity describes Computer as a system that takes actions in docs, apps, emails, and research through hundreds of connectors, allowing it to move beyond simple chat responses into direct manipulation of a user’s workflows. Rather than responding to a single prompt and stopping, Computer deploys sub-agents that work asynchronously in the background. A user can assign a complex workflow, such as monitoring competitor pricing across multiple sources or drafting and sending a weekly report, and the system handles the steps on its own schedule. All of this runs inside a secure isolated cloud sandbox, meaning the processing happens on Perplexity’s infrastructure rather than on a user’s local machine.
The always-on scheduling and monitoring capabilities set Computer apart from conventional chatbot interactions. Traditional AI assistants require a user to stay in the conversation loop, issuing follow-up prompts and reviewing outputs in real time. Computer instead operates more like a background service. Once a task is defined, it can continue executing across sessions and time zones, checking in with the user only when it needs clarification or approval. That asynchronous design is what makes the “months nonstop” framing plausible in theory, even if the practical limits tell a more complicated story about how long those workflows can run before hitting resource constraints or requiring human intervention.
19 Models, One Orchestration Layer
The technical architecture behind Computer is built around parallel model orchestration. According to Perplexity’s changelog, the system coordinates work across 19 models simultaneously, routing different parts of a task to whichever model is best suited for it. A research subtask might go to one model optimized for web retrieval, while a writing subtask routes to another tuned for long-form generation. This multi-model approach means Computer is not locked into the strengths or weaknesses of any single large language model, and can in principle swap components as newer or more specialized models become available.
Perplexity frames this as unifying “every current AI capability into one system.” That is a bold claim, and the company has not published independent benchmarks to back it up. Still, the design philosophy is clear: instead of asking users to pick the right AI tool for each job, Computer absorbs that decision-making into its orchestration layer. For users who currently juggle separate subscriptions to different AI services for coding, writing, data analysis, and search, the pitch is consolidation and reduced cognitive overhead. Whether the 19-model ensemble actually outperforms dedicated single-model tools on specialized tasks is a question Perplexity has not yet answered with public data, leaving early adopters to evaluate performance largely through hands-on experience.
Credits as Fuel, and Their Limits
Computer is available on the web exclusively to Max subscribers, who receive 10,000 monthly credits plus a one-time 20,000-credit bonus when they start using the service. These credits are the currency that powers every action Computer takes, from sending an email to running a multi-step research pipeline. The credit system creates a direct relationship between how much work a user assigns and how quickly they burn through their monthly allotment. Perplexity has not disclosed a public rate card showing how many credits a typical task consumes, which makes it difficult for prospective users to estimate whether 10,000 credits per month will cover their needs or whether they will routinely need to purchase additional capacity.
The fine print introduces real constraints on the “nonstop” promise. Monthly credits do not roll over, so unused credits vanish at the end of each billing cycle instead of accumulating for future large projects. Purchased credits can expire after one year of inactivity, and bonus credits expire on a stated date, meaning that users cannot simply stockpile a large buffer and forget about it. Most notably, tasks pause when credits run low and resume only once the balance is replenished. That pause-and-resume mechanism means a long-running workflow is not truly autonomous. It is tethered to a billing meter that can interrupt it at any point, requiring the user to step back in and either wait for the next monthly refresh or buy more credits to keep the system working.
The Dependency Cycle in Always-On Automation
The tension at the center of Computer’s design is structural. Perplexity is selling the idea of a digital worker that operates independently for extended periods, yet the credit system creates a recurring dependency loop. A user who sets up a monitoring task expected to run for weeks or months must ensure their credit balance stays funded throughout. If they forget to check, or if a complex task burns credits faster than expected because it branches into more subtasks, the work simply stops. The user then has to diagnose where the task paused, confirm nothing was lost or duplicated, and restart it. That is not autonomous work. It is work with a financial kill switch that can trigger without warning if usage estimates are off.
This design choice likely reflects the real cost of running 19 models in parallel on cloud infrastructure. Perplexity cannot offer unlimited compute without a metering system, and credits are the mechanism that aligns user demand with backend costs. But the result is a product that promises freedom from manual task management while quietly requiring a different kind of manual oversight: budget management. For professionals who adopt Computer to handle recurring, long-duration workflows, the credit ceiling becomes the new bottleneck, replacing the human labor it was supposed to eliminate with financial vigilance over a consumption meter. That trade-off may be acceptable for organizations that already track cloud usage closely, but individual users may find that the mental load of monitoring credits undercuts some of the convenience they were seeking.
Where Computer Fits in the AI Agent Race
Perplexity is entering a crowded field. Multiple companies, including Anthropic, Google, and OpenAI, have released or announced AI agent products that aim to handle multi-step tasks with minimal user input. What distinguishes Computer is the breadth of its connector ecosystem and the explicit framing as a persistent background worker rather than a conversational assistant. The hundreds of connectors and the sub-agent architecture suggest Perplexity is targeting users who need integration across many tools, not just a smarter chatbot, positioning Computer as infrastructure for knowledge work rather than a standalone app.
The early access restriction to Max subscribers on the web signals that Perplexity views Computer as a premium offering, not a mass-market feature. The 10,000 monthly credits and initial 20,000-credit boost are generous on paper, but the lack of transparent pricing for different task types makes it hard to compare against rival agents that might bundle usage into flat-rate plans. In practice, Computer may appeal most to power users and teams willing to experiment with orchestrated agents and to live within a metered model in exchange for deep integration and parallel model access. Whether that niche can support Perplexity’s ambition of an “independent digital worker” will depend less on how many models it orchestrates and more on whether users feel that the value of continuous automation outweighs the friction of keeping the meter running.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
