The deep research API with depth you control Research API

A deep research API takes a question and runs an autonomous, multi-step process to answer it: planning a strategy, issuing many searches, reading full pages, cross-referencing sources, and writing a report with citations. It differs from a web search API, which returns ranked results for your own model to process, and from an answer API, which synthesizes a single search pass. Developers use one when the question is too complex for one search and too costly to research by hand.

Every response includes content, content_type and sources. content is a Markdown report with inline citations such as [[1, 2]] by default, or a JSON object when you provide an output_schema. content_type is text or object accordingly. sources lists the web pages the agent read and cited, each with a URL, title and relevant snippets, so every claim can be traced to its evidence.

The Web Search API returns raw results for you to process, in a single fast round trip. The Research API reads, reasons over and synthesizes results for you, across multiple searches and steps, and returns a finished answer. Use Web Search for RAG pipelines, your own search UI or data gathering. Use Research when you want a ready-to-use, cited answer and can trade some latency for depth.

research_effort controls the compute budget. lite ($12 per 1,000, under 10 s) suits quick factual lookups. standard ($50, 10 to 30 s) is the default for most production use. deep ($100, under 120 s) handles complex multi-source synthesis. exhaustive ($450, under 300 s) verifies claims across dozens of sources. frontier ($1,200, background only, p50 about 5 minutes) is for long-horizon investigations. Pick the lowest tier that answers the question.

On DeepSearchQA, Google DeepMind's benchmark for research agents, the You.com Research API frontier tier scores 86.61% accuracy and 94.35% F1, up from 83.67% and 93.16% at launch when it ranked first among publicly reported results. On FRAMES and BrowseComp, the exhaustive tier is in the highest accuracy range at a fraction of the latency of comparable offerings. Full results and methodology are published on the You.com blog.

Yes. Pass an output_schema using the supported JSON Schema subset and the API returns a JSON object that follows it in content, with content_type set to object. This is how teams turn research into typed records for databases, dashboards or downstream agents. Structured output is available on standard and higher tiers, not on lite.

Yes. source_control (beta) lets you constrain the research agent through domain, freshness and country filters, so the evidence comes from sources you trust and from the time window you care about. It is available on the Research API; the Finance Research API uses a dedicated financial index instead and does not accept source_control.

Latency scales with effort: under 10 seconds for lite, 10 to 30 seconds for standard, under 120 seconds for deep and under 300 seconds for exhaustive. A synchronous request holds the connection for the full run, so set your HTTP client timeout above the tier's upper bound. For deep, exhaustive and frontier work, send background: true to queue the request and poll or stream for the result instead.

Setting background: true runs the request asynchronously and returns a task handle. Your workers stay responsive while the research runs, and you collect the result by polling or streaming the task. It is optional for most tiers and required for frontier, whose runs can take from 30 seconds to several hours depending on the question.

Yes. OpenAI retired its o3-deep-research and o4-mini-deep-research models in July 2026, and Perplexity ended Sonar Chat Completions in September 2026. The Research API is a straightforward replacement for that workload: send your question as input (up to 40,000 characters), pick an effort level, and get Markdown with citations or JSON through output_schema. Pricing is fixed per call rather than per token, long runs use background mode with polling or streaming, and the same key works over REST and MCP. Your free credits cover re-running your existing evaluation set before you switch.

Complex question answering such as comparative analyses, due diligence and market research with citations you can trace, internal knowledge assistants that answer product, regulatory or technical questions, and content pipelines where a cited research draft becomes the source material for a report or briefing. Agents also call it as a tool when a task needs a researched answer rather than a list of links.

Yes. The You.com MCP server exposes research as a tool at api.you.com/mcp/research, alongside search and contents at api.you.com/mcp. Claude, Cursor, VS Code and other MCP clients can run research without custom glue code, and the same API key covers REST and MCP.

New accounts get $100 in free credits with no credit card required. That covers roughly 2,000 standard-tier research calls or 8,000 lite calls, enough to evaluate quality on your own questions and ship a proof of concept before paying anything.

Pricing is fixed per tier and billed per call: lite $12, standard $50, deep $100, exhaustive $450 and frontier $1,200 per 1,000 calls. There are no token charges, minimums or platform fees. Volume discounts and annual pricing are available for high-throughput applications.

No. Your questions, reports and sources are never used for model training. Zero Data Retention (ZDR) is available for enterprise agreements and automatically purges request and response content. The platform is SOC 2 certified and DPA-ready.