Perplexity's NEW Computer Just DESTROYED Every AI Tool You're Paying For

High-level summary

Perplexity Computer (running in Perplexity’s Comet browser) is demonstrated via a hands-on review. The presenter gives five real-world jobs to the system, showing prompts, activity logs, and final outputs to highlight how the product orchestrates multiple AI models and connectors, runs multi-step workflows (agents) in the background, and integrates with the apps and data you already use so you don’t have to copy/paste between tools.

The video is a demo-style walkthrough that emphasizes real-world tasks, transparency of agent activity, parallel execution, and extensibility via a skills marketplace.

Key technological concepts and product features

• Comet browser

  • Perplexity’s browser extension/new-tab UI where “Computer” runs.

• Tasks

  • Workspace where each job is saved; multiple tasks can run concurrently.
  • All outputs (reports, images, code, clips) are auto-saved.

• Connectors

  • Native integrations to external tools (Google Drive, Gmail, Slack, GitHub, Salesforce, Snowflake, CB Insights, etc.) allowing the AI to read/write to real data sources.

• MCP (Model Context Protocol)

  • A “universal language” layer that lets models interact with software/tools (analogy: universal remote). Enables connector interoperability.

• Skills

  • Pre-built capabilities (research assistant, media processor, finance markets, web app builder, data visualization, etc.).
  • System auto-selects skills based on the task; users can add third-party skills from a marketplace (skills.sh).

• Orchestrator model selection

  • You can choose which model runs orchestration (example used: Claude Opus 4.6 for complex reasoning).

• Credits system

  • Tasks consume credits (example: a content strategy task cost 97 credits on the Pro plan; Pro ≈ 4,000 credits).

• Activity log / transparency

  • Live step-by-step log showing what agents do (which skills they launched, browsing activity, API calls). Not presented as a black box.

• Parallelism + agents

  • True parallel execution: multiple agent workflows run simultaneously.
  • Agents can self-diagnose and self-fix code errors without human intervention.

• Marketplace & extensibility

  • Install new skills from skills.sh by pasting a URL; installed skills become available to tasks.

• Auto media processing

  • Transcription with timestamps, automated clip selection, video extraction, auto-generated thumbnails and headlines (presenter estimates ~80% useful out of the box).

• Full-stack app generation

  • Demonstrated building a working stock-educator full-stack app with zero user code by combining finance, web app, and data-visualization skills.

• Data-source driven research

  • Can pull premium data (e.g., CB Insights) via connectors, cross-reference web and paid sources, and produce formatted reports.

Concrete demos / tasks performed

1. YouTube content strategy audit

   • Prompt: Analyze the top 5 performing videos + comments and recommend next videos.
   • Process: Orchestrator launched a research assistant, browsed YouTube, read comments, pulled views/links.
   • Output: Content strategy recommendation (70% mastery guides, 30% podcasts); cost = 97 credits.

2. YouTube clip generator + thumbnails

   • Prompt: Analyze a full video, cut clips, and produce thumbnails.
   • Process: Full transcription with timestamps; algorithmic selection of 5–6 clip-worthy segments; video extraction; autogenerated thumbnails and headlines (~80% quality).

3. Full-stack stock educator app

   • Prompt: Build an app to input 3 stocks, visualize 10-year performance, map historical reasons for moves, and teach entry/exit points.
   • Process: Installed an external stock-analysis skill from skills.sh; launched finance, web-app, and data-visualization skills. Encountered an Anthropic SDK import error mid-build; the agent read its own error log, rewrote code, re-ran the build, and finished without human input.
   • Output: Working app URL with 10-year charts, 52-week high, current price, and event-mapped insights. Presenter cautions it’s for learning, not investment advice.

4. Market research across CB Insights + web

   • Prompt: From market insights, identify 10–15 jobs most impacted by AI and the top 5 industries to build in.
   • Process: Used the CB Insights connector and web sources simultaneously, cross-referenced data, and compiled a report.
   • Output: Top industries identified include healthcare, financial services, and education, plus detailed role-level breakdowns and skills in demand.

5. Parallel task throughput demo (implied)

   • Multiple internal tests run in parallel to illustrate throughput and compounding productivity.

Notable behaviors and claims

• Self-debugging agents: Agents can detect errors (e.g., SDK import failure), patch code, and continue without human intervention.
• Real browsing: The system actually navigates external sites (YouTube, CB Insights) rather than hallucinating results.
• Auto skill choice: The system picks which skills to use for a job; users can override the orchestrator model but may not always control skill choice.
• Parallel workflows compound productivity: Running many complex tasks simultaneously is a core differentiator.
• Orchestration layer: Perplexity Computer acts as a control/orchestration layer that runs existing LLMs (Claude, GPT, etc.) and integrates with enterprise tools; it isn’t positioned as a replacement for base models.

Performance and cost examples

• Example: The content strategy task cost 97 credits on the Pro plan (Pro ≈ 4,000 credits).
• The UI shows credit usage in real time during tasks.

Limitations and cautions

• Thumbnails and copywriting are not perfect initially (roughly 80% usable).
• Not a substitute for professional investment advice; stock analysis demo is educational only.
• Technical errors (e.g., SDK imports) can occur, though agents may self-repair during the workflow.

Comparisons and context

• Compared to OpenClaw (referenced): both trend toward local or hybrid model orchestration. Perplexity Computer runs in the cloud, while OpenClaw runs locally.
• Presenter referenced his separate “Cloudbot”/OpenClaw deep-dive videos for additional under-the-hood details.

Guides, tutorials, and resources mentioned

• Demonstrations covered:
  • How to open Computer in the Comet browser and where to click.
  • How to choose the orchestrator model.
  • How to install skills from skills.sh.
  • How to add connectors like CB Insights.
• Prompts used for each task were shared with the presenter’s WhatsApp community and linked in the video description.
• Links promised to related breakdown videos (Cloudbot/OpenClaw) and the creator’s channel subscription.

Main speakers and referenced systems

• Presenter: Vibhav Cicinti (the video host and demonstrator).
• Products / systems: Perplexity Computer, Comet browser, MCP (Model Context Protocol), skills.sh.
• Models & SDKs referenced: Claude Opus 4.6 (Anthropic/Claude), ChatGPT (referenced), Anthropic SDK (error encountered).
• Data/connectors: Google Drive, Gmail, Slack, GitHub, Salesforce, Snowflake, CB Insights.
• Other referenced projects/videos: “Cloudbot” breakdown, OpenClaw.