Perplexity Unveils Perplexity Computer, a Multi-Model Agent System for End-to-End Project Workflows
Original: Introducing Perplexity Computer. View original →
What Perplexity announced on X
Perplexity introduced "Perplexity Computer" on 2026-02-25, describing it as a unified AI system for full project execution. The launch post says the system can handle end-to-end tasks including research, design, coding, deployment, and ongoing project management.
Unlike a single-model assistant framing, Perplexity's thread emphasizes orchestration. The company says Computer can route work across 19 models and assign different subtasks to different model backends in parallel.
Product details shared in the launch thread
- "Massively multi-model" architecture with model routing across 19 models
- Persistent memory, connectors, files, and web access in one workspace
- Usage-based pricing with optional model selection and spend caps
- Max plan users receive 10,000 monthly credits according to Perplexity's post
Perplexity also says the first release is available on web for Max subscribers, with Pro and Enterprise availability planned next.
Why this is high-signal
The launch is notable because it positions agent tooling as an operating layer rather than a single chatbot endpoint. The messaging focuses on workflow continuity: going from one task to many concurrent projects while preserving memory and control over model-level cost routing.
If Perplexity executes this as described, the product would move beyond "ask-answer" interactions into persistent multi-agent task management. The practical differentiator to watch is whether the orchestration claims translate into reliable automation quality across long-running projects.
Primary sources: launch post, multi-model details, availability post.
Related Articles
At its Code with Claude London event, Anthropic launched self-hosted sandboxes (public beta) and MCP tunnels (research preview) for Claude Managed Agents, enabling enterprises to run AI agents entirely within their own infrastructure without exposing sensitive data.
Bumblebee is a read-only scanner for macOS and Linux developer endpoints. Perplexity says it checks risky packages, MCP configs, editor extensions, and browser extensions without invoking package managers or install scripts.
TrapDoor pushed more than 34 malicious packages across npm, PyPI, and Crates.io after May 22. The sharpest twist is not just credential theft, but the attempt to poison .cursorrules and CLAUDE.md files read by AI coding assistants.