The best deep research APIs for developers in 2026

A deep research API takes a question or a task, plans its own searches, reads many sources and returns a cited report or structured data, usually as a job that runs for minutes while you poll or stream it. Which one is best depends on what you need back. For research on specific people and companies, Sixtyfour returns the fields you define, each with its sources and an optional confidence score, and on RECON, a people research benchmark we built and on which five of the systems are ours, its XHigh tier recorded 70.6 percent weighted accuracy against 41.1 percent for the strongest system from another company. For cited reports on any topic, the main options as of November 2026 are Google's Gemini Deep Research agent, Parallel's Task API, Exa Agent, Perplexity's Agent API, Tavily Research, TinyFish's TinyResearch and Valyu DeepResearch, plus OpenAI's models with web search, the path OpenAI points to since its deep research models shut down in July 2026.

We make Sixtyfour, so it is first, and its section sets out the evidence for that placement. Every product and price fact about another company comes from its own pages, read in November 2026. If your agent only needs pages to read, our list of Exa alternatives covers the search APIs one layer down.

We included APIs a developer can call from code that take a question, a task or a record, plan and run their own searches, read the sources, and return a report or structured output with citations. Each one had to have public documentation and pricing we could read in November 2026, and be on offer today.

Deep research APIs do different jobs, and the difference decides which one fits.

1. General deep research. You send a question about any topic, such as a market, a technology or a scientific field, and get a long report with citations. Gemini Deep Research, TinyResearch, Tavily Research and Valyu DeepResearch are built around this job, and most of the others can do it too.
2. Research on a specific person or company. You send what you already know about one subject and the fields you want, and get those fields back with a source for each. Sixtyfour is built for this job, and Parallel's Task API and Exa Agent also return structured fields about entities you describe.
3. Search that returns links. You send a query and get ranked pages or excerpts, and your own agent does the reading. These are search APIs rather than deep research APIs, and most of the companies here sell one alongside their research product.

After Sixtyfour, the order is not a ranking. We know of no public benchmark that has run all nine on the same tasks, and RECON, the one we can speak for, measures research on people only.

Sixtyfour is a deep research API for people and companies, also sold through a web app. People Intelligence takes what you already know about a person in lead_info, any combination of name, title, company, location, LinkedIn URL and email, and the fields you want in struct, each described in plain English. Company Intelligence does the same from a company name or domain, and can also find the people who work there. Both research sources across the web and return your fields in structured_data, with a short research summary and references, a map from each source URL to what was found there. An optional research_plan gives the agents plain-language guidance on where and how to look, up to 100,000 characters.

Turn on confidence scoring with field_confidence: true and each value gets a score from 0 to 100, a short justification and the URLs it was read on, at no extra charge. The docs say each score reflects "how much assumption is needed to tie the value back to the target person or company". Every request can run sync, async with a job to poll, or in bulk from an uploaded file, and anything the agents could not establish is reported as open for a person on your team to decide.

For open-ended work there is Atlas, Sixtyfour's multi-agent investigation product. A team of research agents builds a workspace of notes and an attribution graph that links every finding to its source, then writes a report you can share, and the investigation stays open to continue later. The xhigh tier of People Intelligence runs the same agents and returns your struct fields from the investigation. Atlas is available to Enterprise customers on request. An MCP server at https://mcp.sixtyfour.ai/mcp gives AI assistants People and Company Intelligence as tools, single or in bulk, plus Atlas where it is enabled, at the same price as the matching API calls.

The evidence for first place is RECON, a benchmark of 514 verified facts about 140 people. Sixtyfour built it and offers five of the systems evaluated, and HUD, which the benchmark page names as its external validator, ran it with our dataset and public harness. A GPT-4.1 mini judge grades each answer correct, wrong or missing, and it agreed with human review on 98.5 percent of a 200-field sample. Weighted accuracy is correct answers minus wrong answers, divided by all fields. Sixtyfour XHigh scored 70.6 percent, with 434 correct, 71 wrong and 9 missing, and Sixtyfour Medium, a tier every organisation can use, scored 44.7 percent. The strongest system from another company, Parallel's Ultra 2x processor, scored 41.1 percent. XHigh does not lead on every measure: GPT-6 Astra had the highest precision, right on 90.6 percent of the answers it gave against 85.9 percent for XHigh, while finding 28.0 percent of the facts to XHigh's 84.4 percent. RECON measures research on people, so it says nothing about how these systems write a report on a market or a technology.

As of November 2026, Sixtyfour charges in credits per record, according to its credit costs guide. People Intelligence costs 0.3 credits at micro, 1 at low and 20 at medium, the tiers every organisation can use, and the deeper scout (10 credits), high (40) and xhigh (250 a unit) tiers are enabled on request. Company Intelligence costs 0.3 to 5 credits a record on its self-serve tiers, and bulk jobs cost the same per row.

Choose Sixtyfour when your product needs specific, sourced facts about people or companies in a schema you define, or when one subject needs an investigation a reviewer can check. For a report on a market or a technology, the general research APIs below are built for that job.

Parallel builds web infrastructure for AI agents, and its Task API is its deep research product. You define the task in plain language or JSON, and it returns structured output "complete with citations and confidence levels", in Parallel's words, or a markdown report with inline citations when you ask for a text schema. Each run executes on a processor you choose, from lite, for basic metadata, through pro for exploratory web research, up to ultra8x, which Parallel describes as for "the most difficult deep research" and which can take up to two hours. Parallel's docs say the Task API "is designed for asynchronous, accuracy-critical research", and they point latency-sensitive agent calls to its OpenAI-compatible Responses API instead. Its hosted MCP server adds web search, page fetch and research to clients such as Claude and Cursor.

As of November 2026, according to Parallel's pricing page, Task API runs cost $5 per 1,000 on lite up to $2,400 per 1,000 on ultra8x, with one charge per run however many fields it fills, and failed runs are not billed. ultra2x costs $600 per 1,000 runs. The Responses API costs $10 to $250 per 1,000 requests, depending on the reasoning effort.

On RECON, Parallel's Ultra 2x processor was the strongest system not made by Sixtyfour, at 41.1 percent weighted accuracy, with Ultra 8x at 37.5 percent and Ultra at 27.2 percent.

Choose Parallel when you want structured research on companies, products or any entity you can describe, at a depth you set for each run, or a cited report from the same API.

OpenAI no longer offers a dedicated deep research model through its API. Its o3-deep-research and o4-mini-deep-research models shut down on 23 July 2026, and their model pages now point to a guide on running longer web research. That web search guide recommends GPT-6 Astra with the web_search tool in the Responses API, at high or xhigh reasoning, and background mode for reports that can take several minutes. Setting return_token_budget to unlimited lets a high-effort run inspect many pages, at some cost in latency and price. Responses carry inline citations, and a sources field lists every URL the model consulted.

As of November 2026, OpenAI's pricing page lists GPT-6 Astra at $10 per million input tokens and $50 per million output tokens at standard short-context rates, with web search at $10 per 1,000 calls plus search content tokens billed at the model's rates, so the cost of a run depends on how long the model researches.

RECON ran OpenAI's models through the benchmark's evaluation scripts, which call the Responses API in background mode with web search and code interpreter. GPT-6 Astra, at max reasoning effort, scored 25.1 percent weighted accuracy and had the highest precision of any system tested, 90.6 percent, on the 159 of 514 fields it answered. The benchmark page notes that it left 19 of the 140 people blank without searching, and that these runs alone do not establish why.

Choose OpenAI when you want to build and control your own research loop on OpenAI models, with the model, the reasoning effort and the tools set per request.

Google's Gemini Deep Research agent "autonomously plans, executes, and synthesizes multi-step research tasks" and produces detailed reports with citations. It is in preview and runs only through the Interactions API, in background mode: you start a task, keep its ID, and poll or stream until it finishes. It comes in two versions, Deep Research, built for speed, and Deep Research Max, for the most comprehensive work. By default the agent uses Google Search, URL Context and Code Execution, and it can also call remote MCP servers and search documents you upload. Google's docs list two limits to plan around: no structured output for now, and a maximum research time of 60 minutes, with most tasks finishing within 20.

Pricing is pay-as-you-go on the underlying models and the tools the agent uses. As of November 2026, Google estimates a typical Deep Research task at about $1 to $3 and a Deep Research Max task at about $3 to $7, figures it says are based on preview rates and subject to change.

RECON has not run the Deep Research agent. It ran Gemini 3.1 Pro at high thinking, with Google Search and URL context, a different setup, which scored 13.4 percent weighted accuracy. The benchmark page notes that scores reflect each system's tools and configuration as well as its model.

Choose Gemini Deep Research when you want a long, cited report on a topic with very little code, for the kinds of work Google lists for it, such as market analysis, due diligence and literature reviews.

Exa builds web search for AI agents, and Exa Agent is its deep research product, "an async, usage-based web research agent API" in Exa's words, for list building and deep research. You describe the data you want and the shape it should come back in, and one run fans out many searches, reads the pages behind them, checks each candidate against your criteria and returns a natural-language answer, schema-validated JSON and field-level grounding, the citations behind each field. Runs can query premium data partners through Exa Connect, and Exa's hosted MCP server offers Agent as a tool. For research inside a single call, Deep Search, the research mode of Exa's /search endpoint, plans and refines its searches and can return structured output with field-level citations, in about 4 to 15 seconds for the deep type.

As of November 2026, Exa's pricing page charges Agent runs at a fixed effort from $0.012 (minimal) to $1.00 (xhigh), or metered by usage with a default cap of $5 a run on auto. Deep Search costs $12 per 1,000 requests. The free tier gives $10 of credits that resets on the first of each month.

On RECON, Exa Agent at xhigh effort scored 23.9 percent weighted accuracy.

Choose Exa when you want research and list building on top of Exa's search, with structured output and a citation behind each field.

Perplexity now runs deep research through its Agent API. Support for Sonar Chat Completions ended on 27 September 2026: synchronous and streaming Sonar requests keep working, reformulated as Agent API requests, and asynchronous ones are no longer supported. Perplexity's migration guide suggests the high effort level as the starting point for teams that used sonar-deep-research, and its migration overview recommends xhigh for state-of-the-art deep research, calling it "the highest quality in the lineup". An effort level is a configuration Perplexity manages, bundling a model, search settings, reasoning steps and tools, and it is updated as Perplexity's evaluations show improvements unless you pin a version. Perplexity describes high as expert-level reasoning with exhaustive source coverage and inline citations, xhigh as adding a code sandbox and longer tool-use loops for open-ended tasks, and ultra as for broad research where coverage matters most. Long runs go in background mode and are polled by ID. The Agent API also lets you choose models from OpenAI, Anthropic, Google, xAI and others, returns JSON Schema structured outputs, and adds tools such as a code sandbox, MCP servers and a people_search tool for professional profiles.

As of November 2026, according to Perplexity's pricing page, the Agent API bills tokens at each model's published rates plus tool fees: $2.50 per 1,000 standard web searches and $5 per 1,000 people_search calls.

Choose Perplexity when you want deep research with a managed configuration that improves over time, or the freedom to run the same research loop on models from several labs.

Tavily describes itself as the web layer for AI agents, and its research endpoint "performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detailed research report." You choose the mini model for narrow, well-scoped questions or pro for complex topics that span several subtopics, or leave it on auto. Pass an output_schema and the result follows your JSON Schema, set a citation format, and either stream progress or poll the task by its request ID. Research is also a tool in Tavily's MCP server.

As of November 2026, according to Tavily's credits page, a research request costs between 4 and 110 credits on mini and between 15 and 250 on pro, depending on how much research it does. Every account gets 1,000 free credits a month, and pay-as-you-go credits cost $0.008 each.

Choose Tavily when your agent already uses Tavily search and you want cited reports or structured research on the same key and credits.

TinyFish builds web APIs for AI agents and launched TinyResearch as a product on 5 November 2026. It takes one hard question, investigates it and returns a report "formatted for decisions", in TinyFish's words: a summary at the top, comparisons and key numbers in tables, clear headlines and a citation for every claim. The Research API searches, evaluates sources and writes the cited report, and you can stream its progress or queue a run and fetch it later. TinyFish says a Deep report usually takes 5 to 20 minutes. It is also available through MCP, a CLI and SDKs. In Deep mode, TinyFish reports scores of 51.74 on DRACO, 77.10 on ResearchRubrics and 46.42 on DeepResearch Bench II, from its own runs on those public benchmarks. It works on the public web for now. TinyFish is a Sixtyfour partner.

As of November 2026, a TinyResearch run costs $2, or $3 in Max mode, according to TinyFish's pricing page, and you pay only for reports that finish. New accounts start with $8 in their wallet, and TinySearch and TinyFetch, TinyFish's search and page-fetch APIs, are free within their rate limits.

Choose TinyResearch when you want a decision-ready report at a flat price per run.

Valyu describes itself as the search and research layer for AI doing knowledge work, and its DeepResearch API takes a brief and returns a cited report plus up to ten files in CSV, Excel, PowerPoint, Word or PDF. It plans the work, searches and reads across the web and specialist sources such as filings, papers, clinical trials, patents and market data, and writes the report in the background, in anything from under 30 seconds to 90 minutes depending on the mode. Citations can carry the exact passage from the source, a JSON Schema returns validated fields in place of prose, and the heavy and max modes split a question across up to four researchers working in parallel.

As of November 2026, according to Valyu's pricing page, each task has a fixed price in credits set by its mode, with searches and page reads included: 8 for instant, 25 for fast, 100 for standard, 500 for heavy and 2,500 for max. Pay-as-you-go credits cost $10 for 1,600, which puts a standard task at about 63 cents, and cover web search, news, arXiv and PubMed. Accounts without a plan get 1,000 or 2,000 free credits a month depending on the email used to sign up, and in the first month up to 1,000 of those credits also work on the specialist sources in its standard plans. After that, SEC filings and clinical trials come with the $99 monthly plan, and patents with the $499 plan.

Choose Valyu when your question depends on filings, papers, trials or patents as much as the open web, or when you want the spreadsheet or the deck along with the report.

• RECON is our benchmark. Sixtyfour built it and offers five of its 23 configurations, and the dataset is available on request because it contains personal information.
• RECON measures how well a system finds specific facts about people. The benchmark page says writing a complete research report needs separate evaluation, and that cost and latency are outside its scope.
• Each RECON score comes from a single run in the September to October 2026 evaluation window, and the page notes that source availability and run-to-run variation can affect results.
• TinyFish's scores on DRACO, ResearchRubrics and DeepResearch Bench II are its own runs, and we have not rerun them.

Start from what you need back, and expect to use more than one of these.

• For a cited report on any topic, start with Gemini Deep Research, TinyResearch, Tavily Research or Valyu DeepResearch, and pick Valyu when filings, papers or patents matter. Parallel's Task API and Perplexity's high effort level also write cited reports, and OpenAI suits a team that wants to run its own loop.
• For research inside an agent loop, where an agent waits on the answer, look at Parallel's Responses API, Exa's Deep Search, Valyu's instant mode, Perplexity's Agent API or OpenAI's Responses API with web search. Sixtyfour, Exa, Parallel, Perplexity, Tavily, TinyFish and Valyu all offer research as tools through MCP servers, and our guide to people research over MCP shows Sixtyfour's tools inside an agent.
• For structured research on people and companies at volume, use Sixtyfour's People and Company Intelligence, async or in bulk, with field confidence on, or Atlas when one subject needs a full investigation. Parallel's Task API and Exa Agent return structured fields about other entities you can describe.
• If your agent does its own reading and reasoning, a search API is usually faster and cheaper than any of these.

Whatever you pick, test it on twenty questions whose answers you already know, and count how often it is right, wrong and blank before you trust it with new ones. A confident wrong answer has to be caught before anyone relies on it, which is why RECON subtracts wrong answers and Sixtyfour reports what it could not establish as open.

A deep research API plans its own searches, reads many sources and returns a cited report or structured data. For a report on any topic, Gemini Deep Research, Parallel's Task API, Exa Agent, Perplexity's Agent API, Tavily Research, TinyResearch and Valyu DeepResearch are built for the job, and OpenAI now points research to GPT-6 Astra with web search after retiring its deep research models in July 2026. For research on specific people and companies, with a source behind every field and an optional confidence score, use Sixtyfour: on RECON, the people research benchmark we built, where five of the systems are ours, its XHigh tier recorded 70.6 percent weighted accuracy against 41.1 percent for the strongest system from another company.