September 2025 Summaries

Perplexity describes its Search API as an AI-oriented search system designed to provide fresh, comprehensive, fine-grained information for applications and agents, addressing perceived limitations in older APIs related to cost, latency, stale indexes, and document-level retrieval. Its infrastructure reportedly indexes more than 200 billion URLs using horizontally scalable crawling and storage systems, machine-learning models that prioritize and schedule indexing, and crawler behavior intended to respect robots.txt directives and site availability. The system uses AI-assisted, continuously updated parsing rules to extract and segment web content into meaningful sub-document spans, enabling more precise context selection for language models. Its retrieval pipeline combines lexical and semantic search, filters candidates, and applies increasingly sophisticated ranking models, including cross-encoders, while using product interactions, automated evaluations, and human feedback to improve relevance. Perplexity also introduced the open-source search_evals framework, which evaluates APIs on single-step and agentic research benchmarks, and reports that its API outperformed tested competitors on both quality and latency, with a reported median response time of 358 milliseconds. The company frames AI-native search as an emerging field facing ongoing tradeoffs among scale, freshness, relevance, context quality, and speed.

Perplexity has launched its Search API, giving developers access to the global-scale search infrastructure behind its public answer engine and an index spanning hundreds of billions of webpages. Designed for AI and conventional applications, the API returns structured, fine-grained and ranked content snippets rather than only document-level results, reducing preprocessing and improving downstream relevance. Perplexity says its infrastructure emphasizes accuracy, source-backed information, and real-time freshness through tens of thousands of index updates per second and AI-driven web-content parsing that adapts using evaluations and user-query signals. The company also released a Search SDK, an API platform, documentation, and the open-source search_evals framework for comparing public search APIs on quality and latency. Perplexity positions the offering as a lower-cost, high-performance and easy-to-integrate option for developers, researchers, enterprises, and startups building internet-scale search and AI-agent capabilities.

A system for asynchronous reinforcement learning fine-tuning of the trillion-parameter Kimi-K2 model achieved 1.3-second cross-machine weight updates from 256 BF16 training GPUs to 128 FP8 inference GPUs. It uses one-sided RDMA WRITE operations, allowing training nodes to write directly into inference GPU memory without destination-side notifications, CPU control logic, extra copies, or modifications to the inference engine. A controller gathers metadata once, creates and distributes a static GPU-to-GPU transfer schedule, and signals training GPUs to replay it after each training step. The implementation distributes parameter-transfer tasks across disjoint DeviceMesh groups and pipelines CPU-to-GPU copying, parameter reconstruction and quantization, RDMA transfers, and synchronization to overlap hardware resources while controlling temporary GPU memory through a configurable watermark. By replacing rank-0 gathering and scattering with parallel point-to-point transfers, avoiding repeated scheduling and RPC overhead, and separating transfer functions into testable components, the approach aims to provide faster, more maintainable large-scale model updates.

Perplexity has partnered with 1Password to integrate the password manager’s browser extension directly into its Comet AI browser, marking Comet’s first AI browser partnership focused on credential security. The companies position the integration as a response to the privacy challenges of AI assistants that can perform tasks such as logging into accounts, managing information, and making reservations on users’ behalf. Built-in 1Password support provides encrypted passwords and passkeys, streamlined account authentication, visibility into credential use, and synchronization across devices and browsers. Perplexity says Comet keeps browsing data locally on users’ devices, while 1Password’s end-to-end encryption protects credentials and personal context from reaching Perplexity’s servers. The extension is available to existing 1Password users in Comet at no additional cost, and new users are offered a 25% discount.

Perplexity has introduced Enterprise Max, a premium enterprise tier designed for organizations that want Max-level AI capabilities alongside security, compliance, and administrative controls. Priced at $325 per user per month or $3,250 annually, it offers unlimited Research and Labs queries, access to advanced models such as o3-pro and Opus 4.1 Thinking, Comet Max Assistant integration, expanded file storage limits, and up to 15 enhanced Veo 3 video generations per month. Enterprise Max also makes security features including audit logs, SCIM integration, configurable data retention, and real-time usage analytics available across an organization with as few as one Max user, while maintaining SOC 2 Type II compliance. Organizations can mix Enterprise Pro and Enterprise Max licenses, and Max subscribers will receive early access to future products, premium data sources, and an upcoming Email Assistant.

Perplexity has launched Perplexity for Government, an initiative aimed at expanding secure AI access for U.S. federal agencies and employees. The company says it will automatically apply zero data usage protections to identified government requests, preventing those interactions from being retained for model training, improvement, or evaluation, while also providing access to its most advanced models and features without requiring a contract or paid subscription. Perplexity argues that many public AI tools offer weaker capabilities and limited safeguards for agency data, positioning its approach as secure by default. It also announced Perplexity Enterprise Pro for Government, a customized enterprise platform designed for agency requirements that provides access to Perplexity and third-party AI models, with discussions underway to offer it through government-wide purchasing channels for $0.25 per agency during the first 15 months.