Search as Code replaces Perplexity's fixed retrieval pipeline with a model generated program that runs thousands of parallel retrieval steps and revises itself as results come in.
Perplexity has put its research agent in charge of writing its own search code, replacing the fixed retrieval sequence that powered Deep Research with a program the model generates on the fly. In practice, the model breaks a question into subtasks, then writes the code that runs thousands of parallel retrieval steps inside a sandbox, calling filtering, deduplication, and reranking primitives from Perplexity's Agent Search SDK along the way. The result is a workflow that branches, compares, and refines as it learns, rather than a single linear scrape.
The change, summarized in MarkTechPost's write-up of Perplexity's product update, extends Deep Research into Computer, Perplexity's multi-model orchestrator that the company says launched in late February 2026. Perplexity describes Computer as model-agnostic, routing subtasks across 20+ frontier models, with Opus 4.6 as the core reasoning engine and Google's Gemini positioned as a sub-agent for deep research calls. The "20+ models" tally is Perplexity's own routing list, and "model-agnostic" still means the routing is Perplexity-curated.
The architectural change is Search as Code, and it is the story. Older Deep Research followed a predetermined sequence: search, read, summarize, write. The new stack hands the sequence to the model itself. The model writes code that assembles the search, runs it in a sandbox, and revises the plan as intermediate results come back. Perplexity publishes the surface through both Computer and the Agent API, with endpoints at POST /v1/agent and POST /v1/responses (the latter is OpenAI SDK compatible), and a deep-research preset in the official Python SDK. Developers pay per call on the Agent API; the consumer version is gated to Perplexity Max.
Computer also reads the user's own files alongside the live web. Per Perplexity, the system draws on premium sources including PitchBook, CB Insights, Statista, and census data, with legal data in preview. Deliverables are reports, briefs, decks, dashboards, and live spreadsheets. Each change to a spreadsheet goes through a preview-approve-reject step before it lands, the kind of guardrail that matters when an agent is rewriting financial models on its own.
Starter use cases ship with the product: a five-year cash flow and profit-margin comparison of major AI chip companies in finance, a US-versus-EU data-privacy mapping in legal, a weight-loss-drugs-and-heart-health evidence pull in healthcare, and a reasoning-versus-cost-versus-context-length benchmark of leading models in technology. The shape of the output, a cited research artifact with branching sub-questions and a preview-approve-reject step, is closer to a junior analyst's working file than a chat reply.
The benchmark table is the place to slow down. Perplexity reports first-party gains on three external tests: Humanity's Last Exam from 36.4% to 50.5%, BrowseComp (an OpenAI benchmark) from 40.7% to 83.8%, and DeepSearchQA (a Google DeepMind benchmark) from 81.9% to 85.0%. The BrowseComp jump is the largest, and BrowseComp is also the most agentic of the three, consistent with the architectural story. None of these numbers have been independently replicated in this turn; the MarkTechPost summary, paraphrasing Perplexity, flags that independent verification still matters, and a single aggregator's write-up is not third-party confirmation.
What to watch next is whether the code-shaped workflow holds up outside Perplexity's own evals. The product makes the orchestrator, not any single model, the deliverable. If routing is genuinely Perplexity-curated, then the quality ceiling is set by the curator's taste in frontier models, and the gating is the Agent API and a Perplexity Max subscription. If independent tests confirm the BrowseComp and DeepSearchQA gains, Search as Code becomes a reference design for how research aggregation should look in 2026. Until then, the most accurate read is that Perplexity has shipped the first commercial system where code generation is the native interface for a research workflow, and the rest of the field now has a shape to copy or refute.