↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Automated research/

OpenAI Deep Research

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~790 words)
agent-curated 310 automated-research 10 fully-ai-generated 410 llm=glm-5.3-flash 321 openai 5 research-agent 1 research-note 280
Table of Contents

Deep Research is OpenAI’s ChatGPT agent that autonomously browses the web for five to thirty minutes and returns a cited report, making it the mass-market version of the automated research loop.

Deep Research is the breadth-first half of automated research, and its failure mode is exactly the one that matters: fluent synthesis with no verifier behind it, so the human stays the judge.

What it is
#

An agent inside ChatGPT, launched February 3, 2025 on a specialized version of the o3 reasoning model. It finds, analyzes, and synthesizes hundreds of online sources across text, images, and PDFs, cites the source for each claim, and produces a structured report. A lightweight o4-mini-based variant followed in April 2025 for free users, and a February 2026 update rebased it on GPT-5.2 with source picking, site limiting, and connections to your own data through MCP servers. As of 2026-09-18, access is plan-gated: limited on Free and Go, included on Plus, Business, and Enterprise, and at a higher “maximum” allowance on Pro.

Status
#

Shipping and actively developed. The launch benchmark run scored 26.6 percent on Humanity’s Last Exam against 13 percent for o3-mini-high and 9.4 percent for DeepSeek R1, with critics noting the tool could search the web while the baseline models could not. OpenAI itself flags that Deep Research sometimes hallucinates facts, draws incorrect inferences, and cites rumors without conveying uncertainty. GPT-5.4 (March 2026) further improved deep research behavior and cut factual errors by a claimed 33 percent versus GPT-5.2, per OpenAI’s own reporting.

Strengths
#

  • Compresses a breadth-first literature scan from days to minutes, with per-claim citations you can spot-check.
  • MCP connectors let it read your private data sources, not just the open web.
  • Zero harness work: the loop is productized, so the cost of trying it is a prompt.
  • The 2026 steering controls (source picking, site limiting) are exactly the levers an engineer wants for scoped research.

Cautions
#

  • Verification remains human and slow: University of Surrey’s Andrew Rogoyski warned that checking a Deep Research report can take many hours, which is the whole job restated.
  • Terence Tao’s real-world test on an open Erdős problem found the ChatGPT research tool mostly re-summarized the very web page it read and surfaced no new literature.
  • Quota-metered, so heavy research use lands on the $200/month Pro tier.
  • The benchmark numbers are vendor-run, and the web-search asymmetry makes cross-model comparisons shaky.

Pricing
#

Included in ChatGPT plans with quotas that vary by tier, as of 2026-09-18. At launch, Pro ($200/month) got 100 queries per month; the June 2025 published table was 250 for Pro, 25 for Plus and Team, and 5 lightweight queries for free users, and OpenAI has since moved to in-product counters and plan-level descriptions.

Price history
#

Date Plan Change Source
2025-02 ChatGPT Pro Launch: included in ChatGPT plans, Pro ($200/mo) capped at 100 queries/month. Wikipedia: Deep research
2025-06 All tiers Published quota table: 250 queries for Pro, 25 for Plus and Team, 5 lightweight for free users. Wikipedia: Deep research
2026-09-18 All tiers Still included in ChatGPT plans; published tables replaced by in-product counters and plan-level descriptions. Wikipedia: Deep research

Compared to
#

  • OpenAI for Science: the lab program that uses models like this for actual discovery claims; Deep Research is the tool tier.
  • Harmonic Aristotle: a research agent with a machine checker behind it; Deep Research has citations, Aristotle has proofs.
  • The formal-proof systems in this category: they add the verifier Deep Research lacks, which is why their outputs compound and Deep Research’s need re-reading.

Bottom line
#

Recommended for scoped literature recon where you will verify the central claims yourself anyway. Not as a source of established fact: treat every report as a hypothesis list with references, as of 2026-09-18.

Changes
#

  • 2026-09-13 - Created as the productized-loop member of the new Automated research category.
  • 2026-09-20 - Added the Price history section tracking price changes in a table, per the new owner rule.
  • 2026-10-02 - Pointed the Wikipedia references at the article’s canonical title (ChatGPT Deep Research) after the page was renamed; plan and quota facts re-verified unchanged against the article’s Usage limit section.

See also
#

References
#