Details
- Perplexity introduces Search as Code, a new search architecture where AI agents write Python that talks directly to Perplexity’s search stack.
- The system replaces traditional stepwise tool-calling with programmatic composition of search primitives inside the model.
- Search as Code can fan out multiple queries asynchronously, then deduplicate, filter, join, and rank results before they enter the model’s context.
- The architecture is available in the Perplexity Agent API and is now the default search mechanism in Perplexity Computer.
- In evaluations on deep research benchmarks DSQA, BrowseComp, and HLE, plus wide research benchmarks WideSearch and WANDR, it matches or outperforms all competing systems.
- On DSQA, Search as Code achieves a score of 0.871 versus Anthropic’s 0.815, while operating at nearly half the cost per task.
- On WideSearch, it leads on benchmark score while also running at a lower cost than alternatives.
- Perplexity’s in-house WANDR benchmark, designed to reflect real professional research workloads, shows Search as Code scoring 0.386 versus the next best system at 0.152.
- Perplexity says WANDR is not yet saturated and plans to release this benchmark publicly in the coming weeks.
- Additional technical details and integration guidance are provided in the Perplexity Agent API documentation.
Impact
By turning search orchestration into model-written Python, Perplexity is pushing agentic search toward a more programmable, low-latency paradigm that reduces reliance on brittle tool-calling workflows. The reported gains on deep and wide research benchmarks, combined with lower cost per task, increase pressure on rival AI agent platforms to optimize both retrieval quality and economics.
