Web-grounded research API returning answers plus a search_results source list, with a Sonar Deep Research model billed per search and citation token.
Developers who want retrieval and synthesis in one call rather than wiring a search index to their own model, since the docs describe the Agent API as returning web-grounded answers with built-in citations in a single request. It suits teams already on OpenAI-compatible tooling: there is an OpenAI compatibility guide, a Python SDK and a TypeScript SDK, and the Agent API can run OpenAI, Anthropic, Google and xAI models against the same web search tools. Long-running literature-style jobs map to the Sonar Deep Research model, which is billed on searches and citation tokens instead of a per-request fee.
Anyone who needs the model to hand back a clean citation list should read the docs first: the citations response field is fully deprecated and removed, sources now arrive only in search_results, and the prompt guide tells you to read URLs from the payload rather than from the answer text. Low-volume or bursty workloads will hit the spend-gated rate limits, which start at 1 query per second at Tier 0 and only reach 3 QPS after $50 of cumulative credit purchases, shared across every model and API key in the organisation. The API docs publish no index size, no accuracy or hallucination figure, and no free credit allowance, and seat or Enterprise pricing is not documented there at all.
Pros
Cons
It gives you sources, but you have to collect them yourself. The docs state the citations field is deprecated and removed, and that sources arrive in search_results entries carrying titles, URLs and publication dates. A multi-step run returns several search_results items sharing one citation id space, so you must gather all of them. Inline markers like [1][2] appear only if your prompt asks for them.
Sonar Deep Research has no per-request fee. Perplexity bills it as $2 per million input tokens, $8 per million output tokens, $2 per million citation tokens, $3 per million reasoning tokens, and $5 per thousand search queries. Because the search count and reasoning length vary with the question, the cost of a single report is not fixed and must be measured against your own prompts.
The documentation does not state one. Pricing is described as pay-as-you-go across the Agent, Orchestrator, Search and Embeddings APIs, and rate-limit tiers are keyed to cumulative credit purchases starting at $0 for Tier 0. No trial credit, monthly allowance or free request quota is published in the docs, so treat the API as paid from the first call until Perplexity states otherwise.
Agent API limits are per organisation in queries per second over a rolling one-second window, shared across all models and API keys. Tier 0 allows 1 QPS. Buying $50 of credits lifts you to Tier 1 at 3 QPS, $250 to 8 QPS, $500 to 17 QPS, and $1,000 or more to 33 QPS, which is also the Tier 5 ceiling at $5,000+.
Facts sourced from the vendor’s own site, last verified 2026-10-10.
The Weekly Source
One email a week. New listings, repriced tools, and the ones we would skip - including this one if its pricing changes.
