February 2026 Summaries

Perplexity has been deeply integrated into Samsung Galaxy S26 phones, serving as the primary search and reasoning system for its own assistant and for Bixby, while gaining unprecedented operating-system-level access to Samsung’s native apps, physical controls, and dedicated “Hey Perplexity” wake word. Preloaded on all S26 devices, the assistant can provide sourced web answers, save information to Notes, create reminders and calendar events, and maintain contextual searches within Samsung apps. Bixby uses Perplexity APIs to combine real-time web search with language-model reasoning, and Samsung plans to extend these APIs into Samsung Browser, including agentic browsing capabilities and an optional Perplexity search default. The arrangement supports Samsung’s multi-agent strategy, which reflects reported user reliance on several AI tools for different tasks, and is expected to place Perplexity across hundreds of millions of devices through assistants, browser agents, and search features.

Perplexity has released pplx-embed-v1 and pplx-embed-context-v1, text embedding model families for standard and document-context-aware web-scale retrieval, each available in 0.6B- and 4B-parameter versions. The models use diffusion-based continued pretraining to convert Qwen3 causal language models into bidirectional encoders, followed by multilingual contrastive, contextual, and hard-negative training to improve query-document matching across 30 languages. They generate native INT8 and binary embeddings through quantization-aware training, cutting storage needs by fourfold and 32-fold respectively while retaining most retrieval quality, and they require no instruction prefixes. Perplexity reports that the models lead or match competing systems on benchmarks including MTEB, ConTEB, BERGEN, ToolRet, and internal web-scale retrieval tests, with the 4B contextual model achieving a reported state-of-the-art ConTEB score of 81.96% nDCG@10. The models are available under the MIT License on Hugging Face and through the Perplexity API, with support for several common inference frameworks.

Perplexity has introduced Perplexity Computer, a general-purpose AI system designed to execute complex, long-running workflows by operating software, browsers, filesystems, APIs, and connected services in ways similar to a human worker. Users describe a desired outcome, after which the system divides work into tasks and sub-agents that can conduct research, create documents, process data, write code, generate media, and resolve problems asynchronously in isolated computing environments. The product builds on Perplexity’s prior work in search, deep research, persistent memory, tasks, and its Comet AI-native browser, while emphasizing a model-agnostic approach that combines specialized frontier models rather than relying on one provider. At launch, it uses Opus 4.6 for core reasoning and can assign models including Gemini, Nano Banana, Veo 3.1, Grok, and ChatGPT 5.2 to particular functions, with users able to select models for subtasks. Perplexity presents the platform as an evolution from chat-based answers and individual agents toward an orchestrated AI “computer” capable of coordinating tools and models over extended periods, initially available to Perplexity Max subscribers and planned for Enterprise Max users.

Comet describes a multilayered security strategy for protecting its web-browsing AI assistant from evolving prompt injection and related attacks, emphasizing that overlapping safeguards are necessary because no single defense is sufficient. Before its public launch, the company commissioned Trail of Bits to conduct threat modeling and adversarial testing, which identified vulnerabilities that were addressed through iterative remediation. It later published its four-layer defense architecture and launched bug bounty efforts, then open-sourced BrowseSafe, including a detection model and a benchmark with 14,719 examples spanning multiple attack types, injection strategies, and linguistic styles. The account emphasizes lessons from external testing, continuous threat modeling, transparent research, and industry collaboration, while stating that Comet will continue regular security assessments, disclosure programs, and scaled investments as the assistant handles more sensitive user tasks.

Perplexity has introduced Model Council, a research feature for Max subscribers that runs a single query across three AI models simultaneously, such as Claude Opus 4.6, GPT 5.2, and Gemini 3.0, then uses a synthesizer model to produce a combined response highlighting areas of agreement and disagreement. The feature is designed to address the varying strengths and blind spots of frontier models, which may perform differently across coding, research, creative, and decision-making tasks. Perplexity positions Model Council as useful for investment research, complex personal or strategic choices, creative brainstorming, and fact verification, allowing users to compare multiple reasoning approaches without manually switching models. It is currently available on the web for Perplexity Max users, with mobile support planned for the future.

Perplexity has released DRACO, an open-source benchmark for evaluating deep research agents on authentic, production-derived research tasks rather than narrow synthetic questions. Built from millions of user requests and refined through privacy filtering, contextualization, difficulty screening, and expert review, DRACO contains 100 tasks across ten domains, each assessed with detailed rubrics covering factual accuracy, analytical breadth and depth, presentation, and primary-source citation. The benchmark uses weighted binary criteria, including penalties for hallucinations and unsupported claims, and employs LLM judges whose relative system rankings were validated across multiple models. In Perplexity’s evaluation of four research systems with browser and code-execution tools, its Deep Research product achieved the highest pass rates in every domain, particularly Law and Academic research, led in accuracy, analysis, and citation quality, and recorded the lowest latency. Perplexity says DRACO, its rubrics, judge prompt, and dataset are publicly available, while future versions are intended to support more languages, multi-turn interactions, and additional domains.

Perplexity’s 2025 enterprise usage data, drawn from customers including many Fortune 500 companies, indicates that organizations are increasingly adopting multi-model AI workflows rather than relying on a single provider. While two models represented more than 90% of usage at the start of the year, usage had become more distributed by December, with four models each exceeding 10% share and the leading model accounting for about 23% of queries. Power users and large enterprise accounts were especially likely to experiment with multiple models, with many switching models within a workday to match particular tasks. Claude was the most common choice for programming and software development, while Gemini, GPT, and Claude variants led different specialized areas such as visual arts, financial analysis, debugging, legal work, and medical research. The report also notes that new model releases often create short-lived spikes in adoption before users settle on tools suited to their workflows, suggesting that flexibility and access to multiple models may be increasingly important as capabilities and preferences change quickly.