What is deep research AI — and when is it actually worth using?

Every major AI assistant now has a "deep research" button. OpenAI launched theirs in February 2025, promising cited reports from five to thirty minutes of autonomous browsing; Google's Gemini version claims to "browse hundreds of websites" and ship a multi-page report while you turn your computer off. Press the button, wait, and get back a long, structured, citation-studded document instead of a chat reply.

If you've wondered whether the button is magic or marketing, the honest answer is: neither. Deep research is a genuinely different way of using a model — but it's slower, it costs meaningfully more, and it's the wrong tool for most questions. Here's how it actually works and how to decide when it earns its keep.

What deep research actually does

A normal chat reply — even one with web search turned on — is a single pass: the model runs a search or two, skims what comes back, and answers from its head plus a few snippets. Fast, cheap, and fine for most questions.

Deep research replaces that single pass with a pipeline. The design isn't a trade secret — it's assembled from a few years of published research:

1. Plan. The question gets broken into sub-questions before any searching starts. Stanford's STORM system showed that researching a topic from multiple planned perspectives before writing measurably improves the breadth and organization of the result.
2. Search wide. Each sub-question fires its own searches — often eight or more queries where a chat reply would run one. Anthropic's engineering write-up on their research system found that running these explorations in parallel, each with its own working memory, beat a single sequential researcher by over 90% on complex research tasks.
3. Read the actual pages. Not just the search snippets. The promising results get fetched and read in full, with source quality weighing the pick — official documentation and primary sources beat SEO listicles.
4. Extract facts, not vibes. What survives from each page is a set of specific claims tied to their source. This traces back to OpenAI's WebGPT work, which showed that forcing a model to answer from collected quotes — rather than from memory — makes the output checkable.
5. Check for conflicts. When two sources disagree — and on pricing, benchmarks, and anything moving fast, they usually do — a good pipeline flags the disagreement instead of silently picking a side. A separate verification pass over draft claims, as in Chain-of-Verification, catches errors the first pass missed.
6. Write from the evidence. The final report is composed from that fact table, with citations inline, so every claim traces back to a page you can open yourself.

Plan4 sub-questions, easiest first
Search ×8queries run in parallel
Read 10 pagesofficial docs ranked above SEO posts
Verify2 sources disagree — flagged, not averaged
The report is written only from facts pulled out of those pages — every claim carries its source, and open disagreements are stated plainly.
 10 sources · 2 passes

That last step is the important one. The quality difference between deep research and a long chat reply isn't eloquence — it's that the report is grounded in dozens of pages the model actually read, minutes ago, instead of whatever its training data remembered.

Where it still goes wrong

Deep research reduces hallucination; it doesn't eliminate it. The documented failure modes are worth knowing:

• Confident errors survive. OpenAI's own launch notes acknowledged the agent "occasionally makes factual hallucinations or incorrect inferences" — delivered in the same authoritative tone as everything else.
• Rumors read like facts. A pipeline that reads forum threads and news aggregators can launder speculation into a cited claim, and reports "may not accurately convey uncertainty." A citation proves the model read a page — not that the page was right.
• Verbatim trust is the real risk. AI researcher Andrew Rogoyski warned at launch that properly checking a long AI-written report "can itself take many hours of human work" — which quietly cancels the time you saved. The practical middle ground: spot-check the claims that would change your decision, follow those two or three citations, and treat the rest as a well-organized starting point.

The tell of a trustworthy report is how it handles disagreement. If every source magically agrees, be suspicious — real research on any live topic finds conflicts, and a report that states them ("vendor docs say X, two independent benchmarks found Y") is showing its work.

When to use it — and when not to

Chat + web search · seconds
Quick answer from a search snippet — right most of the time, thin on detail.
1 search · 1 credit
Deep research · minutes
Structured report: options compared, numbers sourced, and the two pages that disagreed called out by name.
8 searches · 10 pages · +10 credits

Deep research earns its cost when the question is wide, factual, and worth being right about:

• Comparing options where the details are scattered across many pages — pricing tiers, feature matrices, hosting alternatives.
• Getting oriented in an unfamiliar topic where you don't yet know the right follow-up questions.
• Anything where you'd otherwise open fifteen tabs and take notes.

Skip it when:

• The question has one answer on one page. A normal reply with web search finds it in seconds.
• You're iterating. Research runs are slow by design; a conversation that needs ten quick back-and-forths wants a fast model, not a pipeline.
• The topic is reasoning, not facts. Refactoring code, drafting an email, thinking through a decision — more sources don't help, a better model does.

The cost difference is real everywhere. ChatGPT meters deep research per month — 25 runs on the $20 tier, 250 on the $200 tier, with lighter-weight runs beyond that. In 1Chat it's simpler: a deep research run costs a flat 10 credits on top of your model's normal per-message rate, the toggle says so before you press it, and it disarms after each send so you never trigger a run by accident.

How 1Chat's version works

We built our deep research on the published research above rather than vibes — the same planning-search-verify structure, with two design choices worth calling out:

• The writer only sees the evidence. The report is written from the extracted fact table, never from raw page dumps — the approach that consistently reduces made-up claims in the research literature.
• Conflicts are surfaced, not averaged. When sources disagree, the report says so and cites both sides.

Every run streams its progress live — the plan, each search, each page read — so you can watch it work instead of staring at a spinner, and the finished report lands in the same chat, with the same memory, as everything else you do.

Turn it on from the + menu in the composer. It costs a lot more than a normal message — 10 credits on top of your model's rate — which is exactly why it's a deliberate toggle and not a default.

Further reading: the Deep Research Agents survey maps the whole landscape of these systems, and DeepResearch Bench is the emerging standard for measuring whether the reports are actually any good.