Perplexity delivered a product-heavy week, deepening its focus on autonomous AI agents and enterprise research tooling. The company integrated its Deep Research capability natively into its Computer agent platform, tying it to search-as-code generation, long-running sandboxes, connectors, tools, and licensed data for Pro and Max users.
Built on Perplexity’s “Search as Code” architecture, the updated Deep Research reportedly executes thousands of retrieval steps in parallel and outperforms the prior version on all benchmarks. Management positions this upgrade as a way to boost product stickiness and support premium pricing, particularly for complex enterprise and professional research workflows.
Perplexity also highlighted Harvard-backed research comparing its Computer agent with traditional search. Across 10,000 matched query pairs over three months, workers were said to complete tasks 87% faster and at 94% lower cost using Computer versus Search, while reporting lower dissatisfaction and higher satisfaction.
The study found Computer usage concentrated in knowledge work such as research, analysis, and document creation, especially among software, finance, and marketing users. Cumulative Computer queries reportedly reached 84 times their first-week level, with sessions averaging about 26 minutes of machine execution and frequent use of external connectors.
Earlier in the week, Perplexity formally unveiled its Search as Code system, which lets AI agents generate Python to call its search stack directly. The approach, now default in Computer and available through the Agent API, is designed to orchestrate async fan-out, deduplication, filtering, joining, and ranking before results hit model context, aiming to cut latency and context pollution.
Company benchmarks suggest Search as Code matches or outperforms competitors like Anthropic on DSQA, leads on WideSearch, and operates at lower cost. Perplexity also promoted its internal WANDR benchmark, intended to mirror real professional research workflows, and plans to release it publicly to standardize evaluation of AI research tools.
On the enterprise front, Perplexity expanded integration with Microsoft 365, bringing its Computer assistant into Excel, Word, PowerPoint, and Outlook via a side panel. These deployments are supported by enterprise-grade controls such as SAML single sign-on, audit logs, and granular admin settings to address security-sensitive customers.
Infrastructure improvements included open-sourcing a rebuilt Unigram tokenizer optimized for XLM-RoBERTa’s 250K-token vocabulary. The company reports 5–6x lower CPU use and roughly 5x lower p50 latency versus common HuggingFace tokenizers, which could reduce costs for reranking and embedding workloads at scale.
Collectively, these updates underscore Perplexity’s strategy of combining technical performance, cost efficiency, and enterprise integrations. If the reported productivity and cost benefits bear out broadly, the company could strengthen its position in AI search, research, and workplace productivity markets, though broader adoption and independent validation will remain key.
Overall, the week highlighted Perplexity’s push to differentiate its AI agent platform through deeper autonomy, benchmarked performance, and tighter alignment with enterprise use cases, potentially laying groundwork for future growth in professional and institutional segments.