OpenAI Deep Research: The AI Agent That Does Your Homework

What Deep Research Actually Does

OpenAI's Deep Research is their second AI agent after Operator, and it solves a specific problem: turning a research question into a comprehensive, cited report without you doing any of the legwork. You type a query, it asks clarifying questions to make sure it understands the scope, and then it disappears for 5 to 30 minutes while it browses the web, reads pages, gathers data, and assembles everything into a structured report.

This is not a chatbot response dressed up as research. The agent plans a multi-step workflow, visits dozens of websites, extracts relevant information from each, and synthesizes it into something that reads like a professional research brief. Every claim gets a citation. Every source is listed at the bottom.

The Model Behind It

Deep Research runs on an optimized version of OpenAI's O3 model, specifically tuned for web browsing and data analysis. At the time of launch, this was the first publicly available model where developers could access the full O3, not just the O3-mini variants that had been released days earlier.

The O3 optimization matters because Deep Research needs to reason about what to search for, evaluate whether the information it found actually answers the question, and decide when to backtrack and try a different approach. This is where the agent behavior shines. Traditional web search tools hit a page, extract some text, and move on. Deep Research reads a page, determines if the content is useful, and adjusts its strategy in real time.

The model also includes a code interpreter. If your research involves data-heavy questions, it can create visualizations, plot charts, and embed them directly in the report. It handles images and PDFs found on web pages too, pulling data from documents the same way a human researcher would.

How the Research Process Works

The workflow follows a clear pattern:

1. You submit a query with optional file attachments for additional context.
2. The agent asks clarifying questions to narrow the scope before it starts.
3. It begins browsing and you can watch the progress as it searches different websites and gathers details step by step.
4. It compiles the report with a sidebar summarizing the research and all cited sources.

The entire process is asynchronous. You submit the request, do something else, and come back when it is done. Reports include tables, formatted sections, reference links, and embedded visualizations when the data calls for it.

One detail from the announcement that stood out: the agent does not just march forward blindly. It backtracks and reacts to real-time information when necessary. If it starts down a path that diverges from the original question, it corrects course. This was a known problem with earlier research tools where you would let them loose, come back 20 minutes later, and find they had wandered off-topic entirely.

Quality Compared to Standard ChatGPT

OpenAI published direct comparisons between GPT-4o and Deep Research given the same prompts. The difference is stark.

Take a UX design query: "Find evidence that shows buttons with icons and labels are more usable than buttons without labels or labels without icons." GPT-4o returns a brief answer with minimal detail. Deep Research returns a multi-page report citing specific user studies, with references at the bottom.

The business research examples follow the same pattern. Ask Deep Research for a market analysis and you get detailed tables, specific metrics, and sourced data points. Ask GPT-4o and you get a competent but surface-level summary.

The gap is not about the underlying model intelligence. It is about time. Deep Research spends minutes reading and cross-referencing sources. A standard ChatGPT response fires back in seconds based on training data. The research agent trades speed for thoroughness.

Benchmarks: Humanity's Last Exam

OpenAI included Deep Research in the Humanity's Last Exam benchmark, which was created specifically because existing benchmarks were becoming saturated with models approaching perfect scores. The test consists of 3,000 questions across 100 subjects, from linguistics to rocket science to ecology.

Deep Research scored 25.3% accuracy. For context, O1 scored 9.1% and O3-mini (high mode) scored 13.3%. The benchmark is intentionally difficult, designed to remain challenging as models improve. The significant gap between Deep Research and the base models suggests that the browsing and extended thinking time genuinely improves the quality of answers on hard questions.

The broader insight: the more time Deep Research spends browsing and reasoning about what it reads, the better it performs. This is the fundamental design tradeoff. It is not optimized for speed. It is optimized for depth and accuracy.

Who This Is For

OpenAI positioned Deep Research for intensive knowledge workers across several domains:

• Finance - market research, competitive analysis, investment due diligence
• Science - literature reviews, methodology comparisons, experiment design research
• Policy - regulatory landscape analysis, impact assessments, cross-jurisdiction comparisons
• Engineering - technology evaluation, architecture research, standards compliance

They also mentioned consumer use cases like shopping research: finding the best appliances, comparing cars, or evaluating furniture options where you would normally spend hours reading reviews and spec sheets.

The common thread is tasks where thoroughness matters more than speed. If you need a quick answer, regular ChatGPT is fine. If you need a researched, cited, comprehensive answer, Deep Research is the better tool.

Pricing and Availability at Launch

At launch, Deep Research was available exclusively to ChatGPT Pro subscribers at $200 per month, with a limit of 100 queries per month. That works out to $2 per research query. Plus and Team users were next in line, with Enterprise after that.

OpenAI mentioned plans for a faster, more cost-effective version powered by a smaller model that would still provide high-quality results, with significantly higher rate limits for all paid users. They also hinted at future integrations with subscription-based and internal data sources, expanding what the agent can access beyond the public web.

At $2 per query, the value calculation is straightforward. If a single Deep Research report saves you 30 minutes to an hour of manual research, and your time is worth more than $2 to $4 per hour, the tool pays for itself. For professionals in finance, law, or consulting where research is a core part of the workflow, the math is obvious.

How Deep Research Compares to Manual Research

The time savings across different disciplines are significant. What would take a human researcher hours of searching, reading, cross-referencing, and writing gets compressed into minutes. The output is not perfect. OpenAI acknowledged that the model can still hallucinate facts, and there may be minor formatting issues. But the baseline quality is high enough that the report serves as a strong first draft rather than something you need to verify from scratch.

The real workflow improvement is not just speed. It is the breadth of coverage. A human researcher gets tired. They check 10 sources, maybe 20 if they are thorough. Deep Research can crawl through dozens of websites, read hundreds of pages, and synthesize it all without fatigue or attention drift.

Consider a practical scenario. You are evaluating three database solutions for a new project. Manual research means opening tabs, reading documentation, searching for comparison posts, checking benchmark data, reading user reviews, and eventually synthesizing it into a recommendation. That process takes 2 to 4 hours if done thoroughly. Deep Research handles the same task in under 30 minutes and produces a formatted report with every source cited.

The output is not a replacement for expert judgment. You still need domain knowledge to evaluate whether the report's conclusions make sense. But it eliminates the most time-consuming part of the process: the gathering and initial synthesis of information from dozens of sources.

Limitations to Keep in Mind

Deep Research is not without constraints. The model can hallucinate facts, especially when sources conflict or when information is sparse. OpenAI was upfront about this at launch.

The 5 to 30 minute wait time is a real tradeoff. If you need quick answers to simple questions, standard ChatGPT is faster and more appropriate. Deep Research is designed for complex queries where thoroughness matters more than speed.

At launch, it was also limited to publicly accessible web content. Internal documents, subscription-based research databases, and private repositories were all out of reach. OpenAI mentioned future plans to expand data source access, but the initial version could only browse what was freely available online.

The 100 queries per month limit on the Pro plan means you need to be intentional about what you send to Deep Research. Burning a query on something you could have answered with a quick web search wastes one of your monthly allocations.

The Competitive Landscape

Deep Research launched into a market where AI-assisted research was already gaining traction. Perplexity had established itself as the default AI search tool. Google was building similar capabilities into Gemini. Various startups were exploring agentic research workflows.

What set Deep Research apart was the depth of output. Perplexity excels at quick, sourced answers to factual questions. Deep Research excels at comprehensive reports that synthesize information across many sources. They serve different needs. A quick factual lookup is a Perplexity query. A thorough market analysis is a Deep Research task.

The use of the O3 model as the reasoning backbone also gave Deep Research a capability advantage over competitors using lighter models. The extended thinking time combined with web browsing created outputs that genuinely resembled professional research reports, not just aggregated search results with citations.

The Bigger Picture

Deep Research represents a specific bet on the future of AI agents: give a model more time to think and act, and the quality of output improves dramatically. This is the opposite of the speed race that dominates most LLM development. While other companies optimize for faster token generation, OpenAI built a product that deliberately takes 5 to 30 minutes to produce a result.

The approach makes sense for knowledge work where accuracy matters more than latency. You do not need your market research report in 2 seconds. You need it to be right. Deep Research trades one for the other, and for the right use cases, that tradeoff is exactly correct.

The broader implication is that AI agents are moving beyond simple question-and-answer interactions. Deep Research is not a chatbot. It is a tool that takes a goal, plans an approach, executes multiple steps, and delivers a finished product. That pattern of goal-oriented, multi-step execution is the foundation of every agent framework being built today. OpenAI just made it accessible to anyone with a ChatGPT subscription.