Deep Research Tools: OpenAI vs Perplexity vs Gemini

The Deep Research Moment

Here's the short answer if you just want a pick. For most people in 2026, Perplexity Pro is the best all-round deep research tool on price and speed, OpenAI and Claude produce the deepest reports for hard or ambiguous work, and Gemini wins when your source material lives in Google or you need a callable API. If you're a developer, Perplexity's Sonar Deep Research and Google's Deep Research API are the two you can build on today, since OpenAI retired its dedicated deep research models in July 2026 and Anthropic never shipped one. The rest of this guide shows how those verdicts hold up across pricing, speed, benchmarks, and real jobs.

The category itself is barely 18 months old. On February 2, 2025, OpenAI announced Deep Research. It was the first agent most people had used that could take a one-sentence prompt, plan a 30-minute investigation, browse dozens of sources on its own, and return with a cited report.

The industry reaction was telling. Within six weeks, Perplexity shipped its own Deep Research (February 14) and opened the Sonar Deep Research API to developers weeks later. Google, which had launched Gemini Deep Research quietly in December 2024, accelerated its rollout and kept upgrading the backbone. Anthropic made Claude's web search generally available on May 7, 2025, packaging the Research feature in the same spring window.

Four labs, one product category, one quarter. That doesn't happen by accident. 2024 was the year context windows crossed 200K tokens, tool use became reliable, and agentic loops stopped silently failing halfway through. Deep research was the first consumer-facing app that made all three feel worth paying for. It's closely tied to the broader shift toward agent protocols we cover in The Agentic Web: Inside the MCP Protocol Wars.

If you write, study, analyze markets, or evaluate products, you're already at a disadvantage if you don't use one. The question is which one, and when.

---

What "Deep Research" Actually Does

It's easy to confuse deep research with chat search. You type a question, you get an answer with links. The mechanics are different.

A chat search (like regular ChatGPT with browsing) runs one or two web queries and synthesizes the top results in seconds. A deep research agent does something closer to what a junior analyst does over an afternoon. It breaks your question into sub-questions, runs dozens or hundreds of searches, reads full pages, follows citations, updates its plan as it learns, and produces a structured report with footnotes.

Ask chat search "what are the main critiques of the Phillips curve?" and you'll get a three-paragraph summary. Ask a deep research agent the same thing and you'll get a 15-page report covering Friedman's natural rate hypothesis, the 1970s stagflation breakdown, rational expectations revisions, post-2008 flattening debates, and recent papers from 2023-2025, each with a source you can click.

The trade-off is time. Runs take between 3 and 45 minutes depending on the tool and the depth. That's the point. You queue one up, work on something else, and come back to a report that would have taken you half a day to assemble manually. For more on restructuring research habits around AI agents, see How to Build an AI-Powered Research Workflow in 2026.

---

Head-to-head: The 4 Consumer Tools Compared

Here's the matrix, with the model backing and pricing verified against current product pages in August 2026.

The one-paragraph profiles:

OpenAI Deep Research is the heavyweight. Runs are slower, reports are the longest, and reasoning is visibly deeper on ambiguous topics. It originally ran on a custom o3 model and now sits on OpenAI's GPT-5.x reasoning line. The 25-per-month cap on Plus is the real constraint. Heavy users burn through it in a week, which is why OpenAI split the old $200 Pro plan into $100 and $200 tiers in 2026.

Perplexity Deep Research is the speed champion. Most runs finish in 2-3 minutes. Reports are shorter and more encyclopedic, ideal for a briefing rather than an essay. It's also the only one of the four whose research agent ships as a proper developer API.

Gemini Deep Research made the loudest gains in 2026. Google upgraded the consumer agent to Gemini 3.1 Pro in April 2026 and, crucially, exposed a Deep Research API so developers can call the same capability. It still pulls from your Gmail, Drive, and Docs alongside the web, and it shows a visible research plan you can edit before the agent runs.

Claude Research is the patient one. Runs regularly hit the 30-45 minute end of the range, and the output reflects it: long-form, nuanced, good at weighing contradictory evidence. Claude's current models (Sonnet 5, Opus 5) carry a 1M-token context at standard pricing, so large source sets don't get truncated.

---

Deep Research APIs: What You Can Actually Call

This is the question the comparison articles keep skipping, and the one developers keep searching for: which deep research agents can you call from your own code, how fast are they, and how good are the citations? The landscape shifted hard in 2026, so here's the current picture.

Three things stand out.

Perplexity is still the only turnkey deep research API. You send a prompt to sonar-deep-research and get back a cited report, no orchestration required. The pricing looks unusual because it's honest about the hidden costs everyone else buries: citation tokens and reasoning tokens are billed separately, and web searches cost $5 per thousand on top of the token bill. If you're searching for the "fastest deep research API," this is the direct answer, because Sonar finishes most runs in under three minutes where the others take five to forty-five.

OpenAI moved deep research into its general model line. The dedicated o3-deep-research and o4-mini-deep-research models were genuinely good, but OpenAI deprecated them in April 2026 and shut them down on July 23, 2026. The replacement path is to run a GPT-5.x model on the Responses API with the web_search tool (and optionally remote MCP servers or file search) in background mode. You get more control and can still reach deep-research behavior, but you assemble the loop yourself rather than calling one model that does it end to end.

Google quietly opened the door in 2026. For most of the category's life, Gemini Deep Research was a consumer feature you couldn't call programmatically. That changed with the Deep Research API through Google's Interactions API, with enterprise support announced for Vertex AI. Combined with Grounding with Google Search (5,000 free searches a month on the Gemini 3.x models, then $14 per 1,000), Google is now a real option for teams that want citations tied to Google's index.

For the "best API for research reports with citations" query specifically, three approaches work. Perplexity gives you citations out of the box and reports strong citation accuracy. Exa and Tavily give you raw, source-attributed search results that you feed into your own model, which is the pattern most production deep research features are actually built on, because you control cost and citation formatting. And any model with a web search tool (Claude, GPT-5.x, Gemini) can cite as it goes if you prompt it to. Enterprises that need audit trails tend to prefer the Exa or Tavily route because every claim traces back to a URL you fetched, not to a model's memory. For why grounding and retrieval quality matter more than raw context size, see Context Rot: Why RAG Still Beats a Giant Context Window.

---

Benchmarks: HLE, BrowseComp, and SimpleQA

Three numbers get quoted most: Humanity's Last Exam, BrowseComp, and SimpleQA. They're useful, and they're also overread.

Humanity's Last Exam (HLE), released by Scale AI and the Center for AI Safety in early 2025, is a roughly 2,500-question multi-domain benchmark at the outer edge of what experts can answer. The scores tell the story of how fast this category moved:

In a little over a year, the best shipped agent more than doubled its HLE score. What HLE measures well is synthesis across domains on genuinely hard questions. What it doesn't measure is whether the agent is good at the work you actually do. Most real research isn't PhD-level physics. It's "summarize recent debates on this topic" or "compare these five products for my use case," where the gap between tools is far smaller than the leaderboard implies.

BrowseComp (OpenAI, April 2025) has become the more relevant standard for these agents. It's 1,266 hard, multi-hop browsing questions that reward persistence and good search strategy. OpenAI Deep Research scored 51.5% at launch, and Google's Deep Research Max reported 85.9% in April 2026. The canonical data point is that a plain browsing model scored under 2% on the same test, so almost all of the gain comes from the agent scaffolding, not the base model.

SimpleQA is Perplexity's strong showing. The benchmark tests short-form factual accuracy, and Sonar Deep Research scored 93.9% (Perplexity, Feb 2025). That's a useful proxy for "does the agent hallucinate facts?", which matters a lot when you're going to cite the output. Note it's an agent-with-search number, so it isn't comparable to the ~50% a bare model scores.

The honest read: benchmarks rank tools reliably at the extreme high end of difficulty and badly below it. The best way to pick is to run the same real prompt through two or three tools on the free tier and compare. Benchmarks are suggestive. Your own test is decisive. For a longer argument about why benchmark obsession misleads, see The AI Thinking Trap.

---

Free Tier Reality Check

The marketing pages all highlight free access. Here's what "free" actually means when you try to use these tools for real work.

OpenAI Deep Research (Free: 5/month). Enough to evaluate, not enough to rely on. A single project often eats 2-3 runs (initial pass, follow-up, clarification). You'll hit the cap by day 10 if you use it for work. Plus at $20/month for 25 runs is the realistic starting tier.

Perplexity Deep Research (Free: roughly 5/day). The most generous of the bunch. Perplexity caps free users at a handful of Deep Research runs a day, which still adds up to more than most people need. Free-tier output is shorter than Pro, and you don't get the newer Sonar variants. For casual use, this is the free tier you actually keep using.

Gemini Deep Research (Free: limited access). Google restructured its plans in 2026, and the cheapest paid tier, Google AI Plus, dropped to $4.99/month, which makes Gemini the low-cost entry point if you don't already pay for AI Pro. Free access exists but runs less often and shorter than the paid tiers.

Claude Research (Pro only, $20/month). No dedicated free tier for the Research feature. The free plan includes chat and web search, but multi-step research is behind Pro. Pro also includes Claude's Sonnet 5 access (Opus 5 on the higher tiers), so the $20 buys you one of the strongest long-context reading models on the market.

If you only pay for one, Perplexity Pro gives you the highest volume-to-price ratio at $20. If you only want the smartest output, ChatGPT Plus at $20 gets you 25 OpenAI Deep Research runs plus everything else in the Plus bundle. For Google Workspace users or the lowest monthly cost, Gemini's AI Plus and AI Pro tiers are the natural pick. Claude Pro makes the most sense if you already use Claude for reading and writing and want one integrated subscription.

---

Which Tool for Which Job

After running hundreds of queries across all four, clear patterns emerge. Here's how I'd route work now.

Academic literature review. Claude Research. The long context window matters when the agent needs to hold 20+ papers in working memory, and Claude is noticeably better at distinguishing between superficially similar claims. Runs take longer, but literature reviews aren't time-sensitive.

Market sizing and competitive intelligence. OpenAI Deep Research. The depth of reasoning on ambiguous strategic questions (why a market grew, what's driving customer switching) comes through clearly here. It's the one I trust most for "help me understand this industry" prompts.

Quick factual briefings. Perplexity Deep Research. If you just need a cited two-page summary before a meeting, Perplexity's 3-minute turnaround is hard to beat. SimpleQA-style factual accuracy is a genuine strength.

Buying decisions and product comparisons. Perplexity or Gemini. Both pull in enough real-world review data (forums, YouTube transcripts, spec sheets) to produce useful side-by-side comparisons. Gemini's advantage is pulling in your own Gmail receipts and Drive notes.

Research involving your own documents. Gemini Deep Research. The Workspace integration is the moat. If you're researching a topic where half the source material is in your Drive (meeting notes, PDFs, old emails), nothing else compares.

Developer integrations and bulk runs. Perplexity Sonar Deep Research for a turnkey agent, or Exa and Tavily if you want to build the loop yourself and control citation formatting. Google's new Deep Research API is the enterprise pick when you want results tied to Google's index.

Synthesizing contradictory evidence. Claude. When sources disagree (e.g., "is fiber actually good for diverticulitis?" or "does the Pomodoro technique work?"), Claude is the most willing to surface the disagreement rather than pick a side prematurely.

One pattern that might surprise people: no single tool dominates. I run the same prompt through two agents for high-stakes work. The cost is about $40/month for two subscriptions, and the benefit is noticeably better output than any single tool produces alone. Chat search and deep research are starting to feel less like competing products and more like a stack you compose.

---

The Missing Piece: Turning Research Reports Into Usable Knowledge

Here's what almost no comparison article mentions. The report the agent produces is not the output of your research. Your understanding is.

A 20-page Claude Research output or a 15-page OpenAI Deep Research report is the start of the work, not the end. Read it once, skim the conclusion, close the tab, and you've paid an agent to summarize something you didn't actually learn. The 2025 MIT Media Lab study on passive AI use (tracked in our analysis of AI's impact on learning) showed that heavy ChatGPT users consistently retained less of what they "read" than active learners did.

The fix is what researchers have done for centuries: annotate. Highlight the claims that matter. Flag the sources you want to verify. Link insights across reports.

This is where Glasp's web highlighter fits into the workflow. Run your research on OpenAI, Perplexity, Gemini, or Claude. Paste the report into a readable page. Highlight directly in the browser as you read. Your highlights sync to your Glasp library, searchable and organized, alongside everything else you've read that month.

A few specific workflows that work:

Highlight, then re-query. Read the report, highlight the 10-15 claims that matter most. Paste those highlights back into the same agent with "dig deeper on these specific points." Iterative rather than one-shot.

Stack reports by topic. When you research the same topic across two tools (say, OpenAI and Claude), highlighting both reports in Glasp lets you see where they converge and diverge. Disagreements are often the most interesting parts.

Use YouTube alongside text. When the best sources are podcasts or talks, YouTube Summary gives you transcript-level summaries with timestamps. Pairing a text deep research report with 3-4 annotated YouTube talks covers a topic more thoroughly than either alone.

Chat with your highlights. Glasp's AI chat can answer questions using your annotations as the source. It's the difference between "what did the agent say about X?" and "what have I actually concluded about X?"

Publish what you learned. The community on Glasp is full of other people researching similar topics. Sharing highlighted reports is a forcing function to finish the research, not just queue more of it. For a step-by-step guide, see How to Annotate Articles the Right Way.

A report you read once is a receipt, not knowledge. The highlight-and-annotate step is what converts agent output into something you actually know.

---

Frequently Asked Questions

Which deep research tool is the most accurate?

On the latest published benchmarks, Google's Gemini Deep Research Max leads Humanity's Last Exam at 54.6% (Google, Apr 2026), a big jump from the 26.6% OpenAI Deep Research posted at the category's launch (OpenAI, Feb 2025). For short-form factual accuracy, Perplexity Sonar scored 93.9% on SimpleQA (Feb 2025), which is excellent. In practical use, accuracy differences between the top tools are smaller than benchmarks suggest. The bigger difference is depth versus speed.

What is the fastest deep research API?

Perplexity's sonar-deep-research is the fastest turnkey option, finishing most runs in under three minutes where OpenAI, Gemini, and Claude take five to forty-five. It's priced at $2 per million input tokens and $8 per million output tokens, plus separate citation and reasoning token charges and $5 per 1,000 searches. If you want raw speed on individual searches to build your own loop, Exa and Tavily return sourced results in seconds.

How long do deep research runs take?

Perplexity finishes most runs in under 3 minutes. Gemini typically runs 5-15 minutes. OpenAI Deep Research takes 5-30 minutes depending on query complexity. Claude Research can stretch 5-45 minutes on hard prompts. If you need an answer now, Perplexity. If you can wait, Claude or OpenAI usually produce more thorough reports.

Can I use deep research tools via API?

Yes, and the options changed in 2026. Perplexity's sonar-deep-research is the only turnkey deep research API: one call returns a cited report. Google now exposes a Deep Research API through its Interactions API and Vertex AI. OpenAI retired its dedicated o3-deep-research and o4-mini-deep-research models on July 23, 2026, so you now run a GPT-5.x model on the Responses API with the web search tool instead. Anthropic offers no dedicated research agent, but you can build one with the Messages API web search tool ($10 per 1,000 searches) and the Claude Agent SDK. Exa and Tavily are the popular "build your own" search-for-AI APIs.

Which API gives the best citations?

For citations out of the box, Perplexity's Sonar Deep Research returns a fully cited report and reports strong citation accuracy. For production systems that need every claim traced to a URL, teams usually build on Exa or Tavily, which return source-attributed search results you feed into your own model, so you control formatting and can show an audit trail. Any model with a web search tool (Claude, GPT-5.x, Gemini) can cite inline if you prompt it to.

Is any deep research tool genuinely free?

Yes, but with limits. OpenAI gives free users 5 Deep Research runs per month. Perplexity gives 5 per day on the free tier, the most generous allowance. Gemini has limited free Deep Research access, and its cheapest paid tier (Google AI Plus) now starts at $4.99/month. Claude doesn't offer Research on its free tier. For casual use, Perplexity Free covers most needs. For regular work, a $20/month Pro plan is the realistic entry point.

How do I stop hallucinations in deep research reports?

Three practical tactics. First, always click at least the top 3-5 cited sources and verify the claim is in the source (hallucinations more often come from mis-citing a real source than inventing a fake one). Second, run the same prompt through a second tool and compare. Disagreements between Claude and OpenAI, for example, are often the places where one of them got something wrong. Third, favor Perplexity for high-stakes factual queries, since its SimpleQA score of 93.9% reflects genuine calibration on short-form facts.

Can deep research tools read my private documents?

Gemini Deep Research has the deepest integration, with native access to your Gmail, Drive, and Docs (with permission). Claude Research supports Google Workspace connectors. OpenAI Deep Research can read files you upload during a session but doesn't integrate directly with cloud storage. Perplexity primarily works against the web. If your source material is largely in Google Workspace, Gemini is the obvious pick.

What's the best way to save and reuse deep research reports?

Export the report as PDF or Markdown, open it in a readable view, and highlight it like you would any long article. Glasp is built for exactly this workflow: highlights sync to a library you can search, link to other highlights, and revisit. Without a highlighting step, most deep research reports get read once and forgotten. This is related to what educators call the "generation effect": information you process actively is retained far better than information you passively receive.

---

Conclusion: The Research Stack, Not the Research Tool

A year and a half after OpenAI's launch, the category has clarified. Deep research agents aren't a winner-take-all market. They're a mix where the right answer depends on what you're researching, how much time you have, where your source material lives, and whether you're clicking a button or calling an API.

If I had to pick one for most knowledge workers in 2026, it's Perplexity Pro. The volume-to-price ratio at $20 is the best of the group, runs are fast enough to fit inside a normal work rhythm, and the SimpleQA accuracy is genuinely strong. For heavier or more ambiguous work, pair it with OpenAI Deep Research or Claude Research. For developers, Perplexity's Sonar Deep Research and Google's new Deep Research API are the two you can build on today.

But the tool choice matters less than what you do with the output. The biggest mistake I see people make is treating a deep research report as finished work. It isn't. It's raw material. The actual knowledge gets built when you highlight the claims that matter, link them to other things you've read, and return to them later when the topic comes up again.

That's the workflow Glasp is designed for. Highlight any report, any article, any YouTube transcript. Build a searchable library of what you actually thought was important. Chat with your highlights later when you need to recall something specific. Share your work with others doing the same research.

The deep research agents will keep getting better, and the APIs will keep multiplying. The ones that don't also get a highlighting layer on top will keep producing reports that get read once and forgotten. Don't build your 2026 research workflow around a single tool. Build it around a stack, and make sure the last link in that stack is the one where your own understanding gets recorded.

Start by running one real research question through two of the four tools this week. Highlight both reports. Compare what you learned. That's the workflow. Everything else is a feature list.