TL;DR
Deep Research is ChatGPT’s slow lane: you ask a question, it spends roughly 5-30 minutes reading dozens to hundreds of sources, and returns a multi-section report with a citation on every key claim. Since the February 2026 upgrade it runs on a GPT-5.2-based research model, asks you to confirm an editable research plan first, and lets you steer the run in real time and restrict it to specific sites. The output looks credible enough that people skip verification — which is the exact step Deep Research exists to make worthwhile. Scope it to one sentence, set source rules in the prompt, then click through 5-8 citations before you ship anything.
What this covers
This guide is for the briefs where the reader will actually check the citations, the numbers, and the logic, and where a regular ChatGPT session is too shallow to hold up. Think investment memos, competitive analysis, policy briefs, due-diligence write-ups, and technical assessments before a buy decision. If your reader will accept “ChatGPT said so,” you don’t need Deep Research — you need a paragraph. If they won’t, Deep Research is the right tier, but only if you scope and verify properly.
What changed in 2026
Deep Research launched in early 2025 on an o3-based model, with a cheaper o4-mini fallback. The February 2026 update moved the main engine to a GPT-5.2-based research model and added the controls that matter for serious work:
- Editable research plan. Before it starts, Deep Research asks clarifying questions and shows a plan. You can edit the plan so the report tracks what you actually need.
- Real-time steering. You can watch active queries, pause, and inject follow-up instructions or new sources mid-run without restarting.
- Site and time-range scoping. Restrict the search to trusted domains or a date window, so it stops citing a random Substack.
- MCP and app connectors. Point it at your own data sources through MCP for internal-plus-web briefs.
- Better output. Full-screen report with a table of contents; export to PDF, DOCX, or Markdown (citations degrade slightly on export — keep the in-app version as the source of truth).
Limits and which plan you need
Deep Research is metered separately from normal chat. These allowances have held steady through June 2026; once you exhaust the standard quota, runs fall back to a cheaper lightweight model automatically rather than blocking you.
| Plan | Price (USD/mo) | Deep Research runs/month | Notes |
|---|---|---|---|
| Free | $0 | 5 (lightweight only) | Enough to try it; US Free has ads since Feb 2026 |
| Go | $8 | Limited | Entry paid tier |
| Plus | $20 | 25, then lightweight fallback | The sweet spot for occasional briefs |
| Pro | $100 / $200 | 250 | $200 tier adds full ~1M-token in-app context |
Pricing and limits are current as of June 2026 and change often; check OpenAI’s plan page before you commit. For heavy weekly research, Plus runs out fast — 25 standard runs is roughly one per workday. If you live in Deep Research, the $100 Pro tier’s 250 runs is the practical floor.
Before you start
- Sharpen the question to one sentence. “Compare X and Y on Z for a [size]-sized buyer” beats “tell me about X.” A vague question can’t focus 30 sources.
- Decide what counts as an acceptable source before the run: peer-reviewed, named industry analyst, vendor doc, or independent benchmark. Write it down — you’ll paste it into the prompt and into the site-scoping filter.
- Block calendar time for verification, not just waiting. The verification pass is roughly as long as the run itself.
- Use the standard run, not the lightweight fallback, for anything that ships. If you’re near your monthly quota, save Deep Research for the brief that needs it.
Step by step
-
Write the scoping prompt as if you’re briefing a careful analyst, not a search engine:
Compare Snowflake and Databricks on cost-of-ownership for a 200-person data team running batch ETL on 5TB monthly. Pull from vendor docs, third-party benchmarks published in the last 18 months, and at least two independent customer case studies. Flag any claim that's only vendor-sourced. -
Review the research plan it proposes. This is the highest-leverage edit you’ll make. If the plan over-weights vendor pages or skips a region, fix it here before the run burns 20 minutes on the wrong scope.
-
Restrict sources if it matters. Add the trusted domains and a date window so the model can’t pad the report with stale or low-quality pages.
-
Start the run and do something else useful for about 20 minutes. Don’t sit and watch — the “did it hang?” reflex doesn’t help. If you spot it drifting in the live progress view, steer it then.
-
When it returns, scan the structure first, not the prose. Are the right sections there? Is a source cited per claim, or only at the end?
-
Read the citation list before the body. If sources cluster on one vendor’s site, the synthesis is biased toward that vendor’s framing.
-
Open and verify 5-8 cited sources at random. For each: does the page exist, and does it actually say what the brief claims?
-
Identify the 2-3 claims most likely to be wrong (specific numbers, head-to-head comparisons, recent dates) and verify those against primary sources, not the cited summary article.
A prompt that produces an honest Deep Research run
Deep Research brief.
Constraints:
- If a claim has no source, label it "unsourced — model inference."
- Do not blend vendor marketing claims with independent benchmarks.
Treat each as a separate evidence class.
- For any number (price, market size, growth rate), cite the
primary source, not a secondary article that quotes it.
- If sources disagree, surface the disagreement instead of
averaging it away.
- End with a "weakest evidence" section listing the 3 claims
least well-supported.The “weakest evidence” closer is the most useful add. It forces the model to self-audit, and the items it lists are the ones you most need to verify by hand.
Quality check
- Click 100% of vendor-comparison citations. These are where fabrication and date drift are worst.
- For every quantitative claim, verify against the primary source. Deep Research over-trusts secondary write-ups, and citation links can degrade on export.
- In a fast-moving domain, treat any citation older than 24 months as a flag — usually it means the search missed current data. The date-range filter prevents most of this.
- Ask: “what would a domain expert have included that this brief skipped?” If you can’t answer, the brief isn’t ready to ship.
How to reuse this workflow
- Keep a
deep-research-template.mdwith your standard scoping prompt structure and the “weakest evidence” closer. - For repeat brief types (quarterly competitive landscape, vendor reviews), save successful runs as a structural template — copy the section headings, not the content.
- Build a per-domain source allowlist (analysts, publications, vendor docs you trust). Paste it into the prompt and load it into the site-scoping filter so the model prefers those.
Common mistakes
- Treating the output as a final draft. It’s a strong first draft and a citation map. Nothing more.
- Skipping the plan review. Editing the proposed plan is the cheapest way to fix scope; skipping it wastes a run.
- Skipping verification because the prose sounds authoritative. The confidence-to-correctness ratio is highest in Deep Research output — exactly the failure mode you’d predict.
- Scoping too broadly. “Tell me about the AI infrastructure market” returns sprawl the model can’t focus 30 sources on.
- Mixing vendor and independent sources without distinguishing them. Half your brief ends up echoing the vendor’s own framing.
- Burning a run when a regular session would do. The waiting cost is real; don’t spend a 20-minute run on a question a 10-minute web search would answer.
- Not budgeting verification time. A 20-minute run with 5 minutes of checking ships worse output than a 5-minute regular session with thorough checks.
FAQ
- How is this different from regular ChatGPT research?: Regular research is one-shot — you steer turn by turn in a normal chat. Deep Research is a metered batch job that reads dozens to hundreds of sources and writes a cited report. Use the slow lane only when shallowness is your actual problem.
- How many Deep Research runs do I get?: As of June 2026: 5 lightweight runs/month on Free, 25/month on Plus, Team, and Enterprise (then a lightweight fallback), and 250/month on Pro. Limits change often — confirm on OpenAI’s plan page.
- What model powers it now?: Since February 2026 the standard runs use a GPT-5.2-based research model; overflow runs fall back to a lighter, cheaper model. The reasoning depth is why a run takes minutes, not seconds.
- Can I trust the citations?: Trust the URL existing more than the claim attached to it. Click through and confirm the source says what’s claimed — citation drift is the main failure mode, and links can break on PDF/DOCX export.
- What if the run takes longer than 30 minutes?: Most finish in 5-30 minutes. If it’s clearly stuck well past that, use real-time steering to narrow it, or cancel and rescope — usually the question was too broad.
- Should I use Deep Research instead of hiring a junior analyst?: For a one-off brief, often yes. For ongoing coverage, no — you want a human building domain knowledge over time, not a fresh run each week.
Related
- ChatGPT for Research — Outline, Source, Synthesize
- ChatGPT Web Search Workflow
- ChatGPT for Research Notes
- Deep Research failed