Deep Research
Why it matters
Deep Dive
A deep research run looks nothing like a normal chat completion. Given a prompt, the system first decomposes the question into a research plan: a set of sub-questions, search queries, and angles worth pursuing. It then loops — issuing searches, opening pages, reading, and deciding what to chase next — for anywhere from a couple of minutes to half an hour, often consulting dozens of sources along the way. Intermediate findings accumulate in working notes, and a final synthesis pass turns those notes into a structured report with inline citations. Under the hood it is closer to the planning loop of an Autonomous Agent than to the single-shot document lookup of classic RAG.
How the Agent Loop Works
The core loop is plan, search, read, reassess. After each batch of pages, the model updates its understanding and picks the next queries based on what is still missing, which sources disagree, and which claims need a second confirmation. This is where it diverges from classic RAG: retrieval is not a one-shot lookup but an iterative process the model steers itself. Browsing happens through Tool Use — search, open-page, and sometimes code-execution tools — and everything the agent has read must fit in the Context Window, so these systems aggressively distill pages into notes rather than hoarding raw text.
Why Reasoning Models Changed the Game
Deep research only became practical once models could sustain long chains of work without losing the thread. A single run may span dozens of search-and-read steps, so the underlying model has to hold a plan, track open questions, and notice when a new source contradicts an earlier assumption. OpenAI's version is powered by a reasoning model fine-tuned with reinforcement learning on real browsing tasks, and competing products lean on their own reasoning-capable flagships. In effect, deep research is Test-Time Compute spent on search: the system trades minutes of Reasoning and browsing for an answer that no single forward pass could produce.
Citations Are Not a Guarantee
The polished report with inline links invites over-trust, and that is the main failure mode. A citation tells you the model opened a page, not that the page says what the report claims, and citation-to-claim mismatches are common enough that spot-checking remains essential. The systems also inherit the biases of the open web: SEO-driven content, outdated pages, and marketing material all get read with the same patience as primary sources, while paywalled or unindexed material stays invisible. Like any generator, the model can still hallucinate details, and it can quietly paper over gaps in the evidence with plausible-sounding synthesis. Treat the output as a strong first draft from a very fast research assistant, not as a finished, fact-checked deliverable.
The Field and How It Is Measured
Gemini shipped the first Deep Research in December 2024; OpenAI's February 2025 launch, initially limited to its highest-priced subscription tier with a monthly quota of runs, is what turned it into a category. Within months Perplexity added one to its answer engine, and xAI released DeepSearch in Grok. Access widened quickly, and by mid-2025 most major assistants offered some form of multi-step research mode. Open-source recreations appeared in parallel, usually wiring a search API to a reasoning model inside an agent scaffold.
Measuring these systems is harder than building them. The most cited early result is Humanity's Last Exam, a benchmark of expert-level questions across many disciplines, where OpenAI's Deep Research answered roughly a quarter of the questions correctly — far above the single-digit scores of standard chat models at the time. Vendors also report strong results on GAIA, an agentic assistant Benchmark. Neither benchmark captures report quality directly, though: citation precision, source diversity, and usefulness to a human reader still get judged mostly by hand, and independent evaluations remain thin.