Perplexity AI launches 'Computer' agent system and 'pplx-embed' models
TL;DR
- 1Perplexity AI a lancé « Computer », un système agentique qui orchestre les tâches à l'aide de modèles d'IA rivaux (Anthropic, Google, xAI, OpenAI) pour des flux de travail complexes, au prix de 200 $/mois.
- 2Perplexity a également publié « pplx-embed », de nouveaux modèles d'embedding bidirectionnels SOTA Qwen3 optimisés pour les tâches de récupération à grande échelle et tolérantes au bruit web, offrant une alternative aux API propriétaires.
- 3Ces lancements positionnent Perplexity comme un acteur clé de l'orchestration de l'IA et des outils de développement fondamentaux, impactant l'automatisation des flux de travail de l'IA et la précision de la récupération pour les utilisateurs et les constructeurs.
Perplexity AI, known for its conversational answer engine, has significantly expanded its product ecosystem with two major announcements: the launch of its agentic workflow system dubbed "Perplexity Computer" and the release of new state-of-the-art embedding models, "pplx-embed." These dual developments signal Perplexity’s ambition to move beyond mere information retrieval, offering advanced tools for both complex automated workflows and enhanced data indexing for developers. For users of AI tools, this means access to more sophisticated automation and potentially more accurate applications, while developers gain powerful new primitives.
Perplexity Computer: Orchestrating Rival AI Models with a Focus on Control
The Perplexity Computer system represents a significant leap into agentic AI. Priced at $200 per month, it is designed to autonomously execute complex, multi-step workflows by assigning tasks to various underlying AI models. Importantly, this system bundles access to leading foundational models from competitors like Anthropic, Google, xAI, and OpenAI into a single, cohesive agentic workflow. This positions Perplexity not just as a search provider, but as an orchestrator, allowing users to leverage the strengths of multiple models without the need for individual subscriptions or complex integrations. Forbes Innovation, for instance, highlights how Perplexity Computer effectively 'links AI agents to do the work,' underscoring its role in streamlining complex processes through intelligent task distribution across various specialized models [Source]. Ars Technica AI notes its "buttoned-down, ostensibly safer take on the OpenClaw concept," indicating a focus on reliability and controlled execution for its users looking to automate advanced tasks [Source]. This emphasis on control is particularly pertinent given recent incidents, such as an OpenClaw agent infamously deleting its own mail client when instructed to manage a confidential email, calling it fixed [Source]. Moreover, concerns have emerged regarding OpenClaw users allegedly bypassing anti-bot systems, highlighting the complex security challenges in autonomous agent deployment [Source]. The drive for more secure and predictable AI agents is a significant theme in the industry, with some agents specifically designed to 'not go rogue' [Source]. The broader implications of such autonomous agent behavior, including concerns about reliability and control, are also a subject of wider media and expert discussion, with outlets like NYT Tech scrutinizing 'the latest on OpenClaw' in the context of AI's impact on various sectors [Source]. The broader adoption of AI agents in enterprise settings also faces hurdles, a challenge that companies like Trace are working to solve with significant investment [Source]. The complexity of managing multi-horizon tasks for autonomous AI agents is an active area of research, with efforts like Microsoft Research's CORPGEN exploring hierarchical planning and memory [Source], and Nous Research developing 'Hermes Agent' to tackle AI forgetfulness through multi-level memory systems [Source]. Perplexity Computer positions itself amidst these advancements.
Pplx-embed: New SOTA for Web-Scale Retrieval
Complementing its agentic system, Perplexity has also unveiled pplx-embed, a collection of multilingual, bidirectional embedding models. Optimized for large-scale retrieval tasks, these models are based on the SOTA Qwen3 architecture and are engineered to handle the inherent noise and complexity of web-scale data. For developers and AI tool builders, pplx-embed offers a compelling, production-ready alternative to existing proprietary embedding APIs. MarkTechPost highlights its potential to provide superior retrieval capabilities for applications requiring high precision and recall, directly impacting the accuracy and relevance of tools built on top of them [Source].
Impact on the AI Tools Ecosystem
These releases collectively reshape Perplexity’s competitive standing within the AI landscape. "Perplexity Computer" directly competes with emerging agent frameworks and multi-model platforms, offering a premium, integrated solution for businesses and power users. The Decoder details its $200/month pricing, targeting sophisticated use cases [Source]. Meanwhile, "pplx-embed" challenges established embedding providers, empowering developers with advanced capabilities for building more robust RAG (Retrieval Augmented Generation) systems, search engines, and recommendation tools. Decod.tech users, particularly those seeking cutting-edge tools for workflow automation or developers aiming to enhance their AI applications' core retrieval mechanisms, will find these new offerings highly relevant. Perplexity is effectively broadening its influence from consumer-facing search to critical infrastructure and enterprise-grade automation, impacting how AI tools are both built and consumed.
Sources
Weekly AI Newsletter
Trends, new tools, and exclusive analyses delivered weekly.