Deep Research guide: autonomous multi-step research in ChatGPT
How Deep Research works, when to use it, how to write a good brief, and how to verify the citations it returns.
Deep Research is ChatGPT's autonomous research agent. It pairs o-series reasoning with web search to produce a cited, structured report in 5–30 minutes.
1. Overview
Deep Research is ChatGPT's multi-step research agent. It browses, reads, synthesizes, and cites, producing a structured report in 5 to 30 minutes. This guide covers the mechanism, ideal use cases, prompt patterns, and citation hygiene.
2. Key points
- Best for market scans, competitor teardowns, literature reviews, tech evaluations — not 'give me a number fast'
- Brief should specify: goal, output structure, citation style, length, audience
- Reports carry footnoted citations; spot-check every load-bearing claim before publishing
- Plus/Team ~25 tasks/month, Pro ~250; limits moved to a 30-day rolling window in 2025; interrupt mid-run with 'expand section 2' or 'rewrite the conclusion'
- Export the finished report to Word / PDF or push it to NotebookLM for a podcast version
3. How it works
The sections below unpack each point. Skim alongside the OpenAI documentation for full context.
4. Practical steps
- Define 'done' up front — what does success look like?
- Pick the right model using the comparison above.
- Run the minimal example end to end; record parameters and the model version.
- Integrate into your existing code.
- Monitor logs and failure modes; review weekly.
5. Common errors and fixes
- 401: invalid or expired API key — rotate it.
- 429: rate limit hit — enable exponential backoff retries.
- 400: bad request — verify the model name and messages shape.
- Poor quality output: try a different model, add few-shot examples, trim the prompt.
6. Next steps
Once you are comfortable here, continue with:
- ChatGPT Complete Guide (2026): From Beginner to Expert
- Prompt Engineering Core Patterns: 8 Templates That 2x GPT Output
- OpenAI API Beginner: Your First GPT-5.6 Call
Key points
- Best for market scans, competitor teardowns, literature reviews, tech evaluations — not 'give me a number fast'
- Brief should specify: goal, output structure, citation style, length, audience
- Reports carry footnoted citations; spot-check every load-bearing claim before publishing
- Plus/Team ~25 tasks/month, Pro ~250; limits moved to a 30-day rolling window in 2025; interrupt mid-run with 'expand section 2' or 'rewrite the conclusion'
- Export the finished report to Word / PDF or push it to NotebookLM for a podcast version
Frequently asked questions
Official references
Related articles
Measuring AI R&D Pace Inside a Frontier Lab: Anthropic's Three Self-Reported Metrics
Anthropic's Sept 17 post proposes three self-reported metrics: an R&D automation index (Claude leads 26% of AI R&D), agent oversight (0.002% of 1B+ decisions blocked), and safety compute (about 6%/12%), with methodology and limitations.
Read articleHumanity's Last Exam Explained: Where 2,500 Frontier Questions Come From and How to Read the Scores
HLE is the most-cited frontier reasoning benchmark: 2,500 cross-disciplinary questions finalized 2025-04-03 after bounty cleanup, published in Nature 2026-01-28, with a dynamic HLE-Rolling variant. Origins, dual scoring, citation rules.
Read articleHow to Read the Benchmark Charts in a Model Launch: Effort Curves, Log Cost Axes, and Safeguard Interventions
Model-launch accuracy-vs-cost charts carry hidden methodology choices. Worked on the 2026-09-01 Claude Fable 5.1 announcement: effort curves, log axes, error bars, scoring conventions, missing cells, safeguard zero-scores.
Read articleSubscribe to GPTMap Weekly
One email every Monday: curated OpenAI updates, deep dives, and best practices. No ads, unsubscribe anytime.
Submitting opens Buttondown in a new tab to confirm your subscription.