Perplexity Launches Computer - a $200/Month Agent Platform That Orchestrates 19 AI Models to Run Projects for Weeks
Perplexity's new Computer product breaks tasks into sub-agents routed across Claude, Gemini, GPT-5.2, and Grok, running autonomously for days or months in isolated cloud sandboxes. Available now for Max subscribers at $200/month.

Perplexity today launched Computer, a cloud-based agent platform that breaks complex projects into subtasks and routes each one to the most suitable model from a pool of 19 - including Claude Opus 4.6, GPT-5.2, Gemini, and Grok. It is available now for Max subscribers at $200/month.
This is not a chatbot upgrade. Computer is designed to run autonomously for hours, weeks, or months - researching, writing code, deploying apps, generating reports, and managing recurring workflows without constant human supervision. CEO Aravind Srinivas called it Perplexity's "next big thing," revealing the team had been building it quietly for two months.
TL;DR
- What: Cloud-based agent platform that orchestrates 19 AI models (Claude Opus 4.6, GPT-5.2, Gemini, Grok, and 15 others) to execute multi-step projects autonomously
- How: Breaks tasks into subtasks, spawns specialized sub-agents, routes each to the best model, runs in isolated sandboxes with real file systems and browser access
- Integrations: Gmail, Slack, Notion, Google Drive, GitHub, Jira, Linear, Confluence, FactSet, and hundreds more via MCP
- Pricing: $200/month (Max tier), 10,000 monthly credits + 20,000 bonus credits at launch
- Available: Now, web only. Pro ($20/month) and Enterprise tiers coming later
- Safety: Isolated sandboxes, scoped credentials, human-in-the-loop gates for irreversible actions, BrowseSafe detection model (though researchers found gaps)
How It Works
Computer functions as an orchestrator. You describe an end goal - "Build me a personal finance dashboard that tracks Indian mutual funds, pulls live NAVs, shows portfolio performance, and emails weekly summaries" - and the system decomposes it into a dependency graph of subtasks. Each subtask gets a dedicated, isolated execution environment with access to a real file system, a real browser, command-line tools, and hundreds of third-party connectors.
The key architectural choice: instead of locking users into a single model, Computer routes each subtask to whichever model is best suited for it. Of the 19 models in the pool, Perplexity has named six:
| Model | Provider | Role |
|---|---|---|
| Claude Opus 4.6 | Anthropic | Reasoning, orchestration, coding |
| GPT-5.2 | OpenAI | Long-context recall, web search |
| Gemini | Deep research, sub-agent creation | |
| Grok | xAI | Fast lightweight operations |
| Nano Banana | Image generation | |
| Veo 3.1 | Video generation |
Users can override the router and pin specific subtasks to specific models - useful when you want sensitive code written by a particular model or need to control token spend. The system surfaces checkpoints for human review only when it hits a decision point or an irreversible action.
What It Can Actually Do
The demos Perplexity showcased go well beyond "answer my question":
Full-stack app development. Describe the app you want. Computer researches available APIs, designs the architecture, writes the code, deploys it, and sets up recurring data pipelines. The finance dashboard example above was a real demo.
Long-running monitoring. "Track every major Mumbai real estate project launch, price trends, and competitor pricing - deliver monthly PDF reports with charts." Computer runs this in the background indefinitely, updating and delivering reports on schedule.
Content production. Analyze competitors, generate content strategies, write scripts and captions for social media campaigns - all coordinated across multiple sub-agents.
Automated job applications. Weekly scanning of job postings, with tailored resumes and cover letters generated per listing.
Perplexity says its employees have been dogfooding Computer since January. One team used it to build a 4,000-row spreadsheet overnight that would normally take a week. Others have published engineering documentation, built internal dashboards, and created analysis pipelines.
A live task stream at perplexity.ai/computer/live shows real projects running in real time.
Pricing
Computer is available now for Max subscribers at $200/month. Each account gets 10,000 monthly credits plus a one-time 20,000 bonus credit grant (expires 30 days after activation). Credits are consumed per subtask, with costs varying by which model handles the work.
Perplexity plans to bring Computer to Pro subscribers ($20/month) after load testing, and an Enterprise tier is in development. For now, the $200 price point puts it directly alongside OpenAI's ChatGPT Pro.
The Safety Model
Every Computer task runs in an isolated cloud sandbox - a walled garden with its own file system and network, fenced off from the user's local machine and primary data stores. This is the opposite of local-first agents like OpenClaw, which run directly on your computer with full system access.
Additional safety layers include scoped credentials (Computer requests minimal-access tokens rather than full OAuth grants), human-in-the-loop gates before irreversible actions (publishing, pushing code, sending emails), read-only defaults, rate limits, and auditable logs for every tool call and file write.
Perplexity also open-sourced BrowseSafe, a fine-tuned detection model that scans web page HTML for prompt injection attacks targeting the agent. The model achieves a 90.4% F1 score on Perplexity's benchmark. However, independent researchers found it incorrectly classified 36% of simple malicious attacks as safe - a significant gap for a safety-critical system. Separately, Brave's security team found indirect prompt injection vulnerabilities in Perplexity's Comet browser that could let attackers access user data through malicious web page content.
The cloud-first approach means Computer cannot control your local desktop, which limits what it can do but also limits what can go wrong. Whether the sandbox isolation holds up under adversarial conditions is an open question.
Competitive Landscape
Computer enters a crowded field. OpenAI's Operator automates browser-based tasks through GUI interaction at the same $200 price point. Anthropic offers Claude computer use at the API level and the desktop-focused Claude Cowork. OpenClaw takes a local-first approach with direct device access. Google's Project Mariner does browser automation through Gemini.
Perplexity's differentiator is the multi-model orchestration. Instead of being locked into one provider's models, Computer picks the best tool for each subtask from across providers. The trade-off is clear: you get model diversity and cloud safety, but you lose the ability to interact with local applications and files.
The bigger question is whether an agent platform that depends on other companies' models has a moat. If Claude, GPT, and Gemini are commodities that any orchestrator can call, what stops the next startup from building the same thing? Perplexity is betting that the routing intelligence, the integration layer, and the workflow execution engine are the defensible parts - not the models themselves.
Sources:
- Perplexity Computer Bundles Rival AI Models Into One Agentic Workflow System for $200 a Month - The Decoder
- Perplexity's New Tool Deploys Teams of AI Agents to Do Your Work - PCWorld
- Perplexity Introduces Unified AI Platform for Research, Coding, Deployment - Benzinga
- Perplexity Computer Orchestrates 19 AI Models to Execute Month-Long Workflows - TrendingTopics
- BrowseSafe: Protecting AI Agents from Malicious Web Content - Perplexity Research
- Comet Prompt Injection Vulnerability - Brave Security
Last updated
