Perplexity Computer: The AI Orchestra Conductor That Runs Projects for $200/Month

On February 25, 2026, Perplexity CEO Aravind Srinivas unveiled Perplexity Computer — an AI agent built not around a single model but around multi-model orchestration of 19 frontier models simultaneously — and opened with a conductor’s line: “Musicians play their instruments. I play the orchestra.” The metaphor is deliberate, and so is the timing: Anthropic launched Claude Cowork on January 14. OpenClaw crossed 219,000 GitHub stars by early February. Perplexity just stepped onto the same stage with a baton.

This is not another chatbot wrapper. Computer decomposes a user’s goal into subtasks, assigns each to the model best suited for it, and runs the workflow autonomously for hours or months without re-prompting. The pitch: stop choosing between the best models. Use all of them.

What Perplexity Computer Actually Does

The architecture centers on Claude Opus 4.6 as the reasoning engine. It receives a user goal, breaks it into discrete subtasks, spins up specialized sub-agents for each, and routes work to whichever model handles that task best. Claude handles reasoning and coding. Gemini covers deep research. GPT-5.2 handles long-context recall and expansive web search. Grok takes speed-sensitive lightweight tasks. Nano Banana handles image generation. Veo 3.1 handles video. The remaining 13 models in the roster have not been publicly named.

The platform connects to hundreds of applications at launch — Slack, GitHub, Google Drive, Jira, Notion, Gmail, Google Calendar, Linear, Outlook, OneDrive, Dropbox, and Box — plus Model Context Protocol support for custom integrations. Computer can browse the web in real time, fill forms, read and write files, execute code, and produce charts and applications. It maintains persistent memory across sessions, so a workflow started on February 25 can pick up exactly where it left off in March.

The safety design is pointed. Every operation runs in an isolated sandbox with scoped credentials — not broad account access. Before any irreversible action, Computer pauses for human review. Srinivas framed this directly as a contrast to OpenClaw, whose agentic AI security vulnerabilities led to a documented near-wipe of a Meta researcher’s email inbox. Perplexity is betting that enterprise buyers will pay a premium for a managed safety guarantee that open-source cannot credibly offer.

Srinivas put the vision plainly: “Perplexity Computer is what a personal computer in 2026 should be. It’s personal to you, remembers your past work, and is secure by default.”

The Orchestration Layer Is Now a Product

Here is the bet Perplexity is making: the routing layer beats the model layer. GPT-5.2, Claude Opus 4.6, and Gemini are all capable enough that marginal quality differences matter less than which platform deploys them most effectively. The orchestration strategy inverts the typical AI competition — rather than racing to build a better model, Perplexity is building infrastructure that makes individual model quality irrelevant to the end user.

The bet has a built-in vulnerability. As Semafor reported, if frontier models become commoditized, Perplexity’s ability to swap between them loses its competitive edge — because every platform could do exactly the same thing. The orchestration advantage depends on orchestration remaining hard.

The model lineup can shift as better options emerge. If a new video model outperforms Veo 3.1 in six months, Perplexity swaps it in. Users never see the change; they just get better output. That is a structurally different relationship with AI progress than a platform locked to a single vendor’s release cycle.

The analogy to cloud computing is apt. AWS won by abstracting hardware complexity so developers could focus on applications. Early hands-on reviews suggest Computer applies the same logic: abstract model selection so users focus on goals, not toolchains. Compare that to Claude Cowork’s enterprise plugin ecosystem, which deepens the relationship between a single model and enterprise workflows rather than routing across multiple providers.

The Pricing Signal: Per-Token Consumer Billing Arrives

Computer launches exclusively for Max subscribers at $200/month (or $2,000/year). Max subscribers receive 10,000 credits per month, plus a 20,000-credit launch bonus valid for 30 days. Pro ($20/month), Enterprise Pro ($40/month per seat), and Enterprise Max ($325/month per seat) rollouts are “coming in the next few weeks,” per Perplexity’s announcement.

This matters beyond the price tag. It is Perplexity’s first consumer per-token billing model. Every previous subscription was flat-rate — pay your monthly fee and ask as many questions as you want. Credits introduce usage-based economics to the consumer AI market, standard in enterprise APIs for years but deliberately avoided at the consumer level to reduce friction.

Analysis of the credit structure raises a practical concern: autonomous month-long workflows could consume credits far faster than a casual user expects. Running Computer on a months-long research project is not the same as running a search query. One early reaction circulating on social media put it bluntly: “A month of computing would cost $1,500.” That may be hyperbole, but the underlying concern is real — per-token billing transfers cost unpredictability from Perplexity’s infrastructure team to the user’s credit balance.

Three Visions of Agentic AI Competing Simultaneously

The agentic AI market has fractured into three distinct philosophies, and they are now in direct competition. Understanding Computer requires understanding what it is not.

Perplexity Computer is a managed, cloud-based multi-model service. The complexity of model selection is invisible to users. Safety is guaranteed by the platform. The cost is $200/month. The trade-off: you run in Perplexity’s walled garden, on Perplexity’s infrastructure, under Perplexity’s data policies.

Claude Cowork (Anthropic) is a single-model enterprise agent with a plugin ecosystem. Its legal plugin triggered a $285 billion selloff in enterprise software stocks on February 3, 2026, and the platform bets on depth over breadth — one exceptional model with deep workflow integrations. OpenClaw’s chaotic open-source debut stands at the opposite extreme: local-first, free, and documented by at least one incident in which it nearly deleted an inbox. 219,000 GitHub stars. Zero managed safety guarantees.

PYMNTS described Perplexity’s positioning as the “managed middle ground” — more control than OpenClaw’s chaos, more flexibility than Cowork’s single-model lock-in. SitePoint’s technical comparison highlighted Computer’s sandbox isolation as the key differentiator from OpenClaw’s local execution model, where a compromised agent has direct access to a user’s actual file system. That is a meaningful safety gap, and Perplexity is charging for the privilege of not having to worry about it.

The Dependency Paradox: Perplexity’s Fragile Supply Chain

Here is the structural risk Perplexity does not advertise. Computer’s value proposition depends entirely on continued access to models owned by its direct competitors. Claude Opus 4.6 is Anthropic’s. GPT-5.2 is OpenAI’s. Gemini and Veo 3.1 are Google’s. Grok is xAI’s. Any one of them could restrict API access, raise prices, or introduce terms that make third-party orchestration commercially unviable. A January 2026 Microsoft Foundry partnership provides some buffer, but it does not eliminate the risk.

The copyright litigation compounds the picture. Active lawsuits from the Chicago Tribune, Nikkei, Encyclopedia Britannica, and Reddit allege Perplexity’s search layer systematically violated robots.txt. If courts restrict PerplexityBot, the real-time research capabilities Computer depends on could be materially constrained. Perplexity’s $20 billion valuation rests partly on the assumption that its search moat survives intact.

Srinivas captured the product’s ambition in his announcement: “The computer is one of the best inventions known to mankind. And when AIs can orchestrate a file system with CLI tools + a browser… AI essentially becomes the Computer, running things on the cloud as you sleep.” That vision is real. But the orchestra metaphor cuts both ways — an orchestra is only as good as its instruments, and Perplexity does not own a single one.

The Bottom Line

Perplexity Computer is a credible entrant in the agentic AI race. The competition is no longer about which company has the best model — it is about who has the best routing layer. Perplexity’s answer is a managed cloud service that abstracts model complexity, provides safety guarantees, and swaps in better models as they arrive. That is a real product with a real target market.

The risks are equally real. Per-token consumer billing will produce sticker shock for users who treat Computer like a search engine. The supply chain dependency on competitor models is structurally awkward for a $20 billion company. Copyright litigation introduces uncertainty that valuation cannot paper over.

The immediate tests are concrete: Pro and Enterprise rollouts in the “next few weeks” will show whether Computer’s value proposition scales beyond early adopters willing to spend $200/month. Usage data — which Perplexity has not published — will show whether autonomous long-running workflows actually work in production or blow through credits on tasks users could have done manually. But the structural question is longer-range: Anthropic, OpenAI, and Google are all currently subsidizing a competitor. When one of them recalculates that arrangement, the orchestra loses an instrument it cannot replace.