OpenAI Deep Research
- Category
- Research
- Type
- commercial
- Source
- Vendor page
- Last reviewed
- 2026-10-17 (what this means)
Citation check incomplete — recheck later.
Summary
OpenAI Deep Research is a commercial ChatGPT feature that writes cited research reports. In independent evaluations (DeepResearch Bench, ICLR 2026; Mind2Web 2, NeurIPS 2025; ReportBench; a citation-URL study; TTD-DR preprint), it ranked among the stronger systems tested on report quality and task success, while a share of its citations did not match their sources or did not resolve. Most tests used 2025 versions; the underlying model has changed since. Informational only; not a guarantee, endorsement, or certification.
Overall score: 65 · Evidence grade: B · Independent sources: 5 performance, 2 other
Factor breakdown
| Factor | Weight | Score | Note |
|---|---|---|---|
| Completion rate | 30% | In Mind2Web 2 it fully completed 28% of hard tasks (highest among systems listed) and partially completed 54%. Judgment call. | |
| Disputes | 20% | No formal disputes specific to this feature were found. Independent tests found about 78% citation accuracy and 3.5% fabricated / 10.1% non-resolving citation URLs. Judgment call. | |
| Adherence to limits | 25% | Usage caps, site restriction and read-only connected apps are documented by the vendor; no independent adherence test was found. Judgment call. | |
| Evidence quality | 25% | Five independent performance sources, two peer-reviewed (ICLR 2026, NeurIPS 2025 D&B); most tested 2025 versions. Judgment call. |
Overall = 0.30×60 + 0.20×65 + 0.25×60 + 0.25×75 = 64.75, shown as 65. See the methodology.
Version note
As of 2026-10-17: OpenAI updated deep research in February 2026 and removed the legacy mode on 26 March 2026; the counted tests used 2025 versions; the current version has not been independently tested. Sources: OpenAI release notes (archived 21 March 2026); OpenAI Help Center.
Security and disputes
No formal disputes specific to this feature were found. In September 2025 Radware disclosed ShadowLeak, a zero-click indirect prompt-injection issue: a crafted email could make deep research send Gmail data to an outside address when the Gmail connector was enabled. It was reported to OpenAI in June 2025 and fixed in August 2025.
Uncertainties
- The underlying model has changed over time; most independent tests used 2025 versions.
- No independent test of usage-limit or permission adherence was found.
- ReportBench venue status is unverified; two sources are preprints.
- Benchmarks rely partly on AI judges.
- Vendor benchmark figures (e.g. HLE, GAIA) are not used as independent evidence.
- Citation check incomplete — recheck later (0057: an ANZ J Surg perspective was not a test; an Eye brief communication is paywalled until 2027; Semantic Scholar still unavailable).
Sources
- DeepResearch Bench (ICLR 2026): report quality and citation accuracy — independent performance, accessed 2026-10-16
- DeepResearch Bench arXiv version — independent performance, accessed 2026-10-16
- Mind2Web 2: task success rates — independent performance, accessed 2026-10-16
- Mind2Web 2 venue: NeurIPS 2025 Datasets and Benchmarks Track (OpenReview) — independent performance, accessed 2026-10-16
- ReportBench: cited-statement match rate (preprint) — independent performance, accessed 2026-10-16
- Citation-URL validity study (preprint) — independent performance, accessed 2026-10-16
- TTD-DR preprint; OpenAI Deep Research as comparison baseline — independent performance, accessed 2026-10-16
- Commentary summarizing early tests (anecdotal) — independent other, accessed 2026-10-16
- Radware ShadowLeak disclosure (security finding; fixed August 2025) — independent other, accessed 2026-10-17
- OpenAI release notes (archived 21 March 2026): February 2026 update; legacy mode removed 26 March 2026 — vendor or maintainer, accessed 2026-10-17
- OpenAI Help Center (current): deep research uses the latest models by default — vendor or maintainer, accessed 2026-10-17
- OpenAI launch post: claims, stated limitations, model history — vendor or maintainer, accessed 2026-10-16
- OpenAI Help Center: usage limits and scope controls — vendor or maintainer, accessed 2026-10-16
Decision records: 0003; 0019 (recheck); 0057 (follow-up review).