Editor’s update, October 19, 2026: This article revisits OpenAI’s February 2, 2025 Deep Research launch. The original Pro-only rollout, 100-query allowance, and promised future features are historical. Deep research remains a ChatGPT feature documented by OpenAI; its availability and usage depend on your account. Separately, the older o3-deep-research and o4-mini-deep-research API models reached their published shutdown date on July 23, 2026. API model retirement does not mean the ChatGPT feature was discontinued.
Image: Official OpenAI Deep Research announcement dated February 2, 2025, captured October 19, 2026. This historical header identifies the original ChatGPT launch; it does not depict a research result, establish today’s usage allowance or imply that the separate API models remain available.
OpenAI introduced deep research in ChatGPT as a way to delegate multi-step internet investigation. Rather than returning only a quick response, the launch feature could search, analyze, and combine online information into a longer report with citations.
The February 2025 version used an o3-derived model optimized for browsing and data analysis. That describes the launch system, not a permanent commitment to the same underlying model. OpenAI positioned it for intensive knowledge work and carefully researched purchases, including science, policy, engineering, and product comparisons.
What Deep Research offered at launch
The original workflow began with a research question and optional uploaded files or spreadsheets. The system planned a sequence of searches, changed direction as it found evidence, and displayed research activity and sources in a sidebar. The intended output was a documented report rather than an unsupported short answer.
- Multi-step investigation: find sources, follow leads, and synthesize information across pages.
- Additional context: use uploaded documents and spreadsheets alongside web research.
- Visible progress: show research steps and sources during the task.
- Cited reports: provide references that readers could inspect and check.
OpenAI said launch tasks could take approximately 5–30 minutes. That was a description of the original experience, not a guaranteed completion time. The useful distinction remains depth versus speed: an extended research report serves a different purpose from a quick factual search.
The original benchmark results, with their conditions
The following figures are OpenAI’s February 2025 launch results. They are preserved as historical evaluation data, not presented as a current leaderboard or independent testing by this website.
Humanity’s Last Exam
OpenAI reported 26.6% accuracy for the model powering deep research, using browsing and Python tools. The announcement compared it with these contemporary results:
| Model | Accuracy |
|---|---|
| GPT-4o | 3.3% |
| Grok-2 | 3.8% |
| Claude 3.5 Sonnet | 4.3% |
| Gemini Thinking | 6.2% |
| OpenAI o1 | 9.1% |
| DeepSeek-R1* | 9.4% |
| OpenAI o3-mini (high)* | 13.0% |
| OpenAI deep research** | 26.6% |
*The announcement evaluated these non-multimodal models on the text-only subset. **Deep research had browsing and Python access. Different tool and modality conditions limit direct comparisons. These percentages are not the probability that any research report will be fully accurate.
GAIA
The launch also reported results on GAIA, a benchmark requiring combinations of reasoning, web browsing, multimodal understanding, and tool use:
| Level | Previous best reported | Deep research pass@1 | Deep research consensus@64 |
|---|---|---|---|
| Level 1 | 67.92% | 74.29% | 78.66% |
| Level 2 | 67.44% | 69.06% | 73.21% |
| Level 3 | 42.31% | 47.6% | 58.03% |
| Average | 63.64% | 67.36% | 72.57% |
Pass@1 and consensus@64 are different evaluation settings; a 64-sample consensus score should not be described as a single-attempt result. OpenAI also said it blocked certain sources because benchmark ground-truth answers had leaked online. The original announcement contains the evaluation notes. These tests do not establish that every real research task is faster or correct.
Limitations that matter when reading a report
The launch acknowledged factual hallucinations, incorrect inferences, difficulty separating authoritative sources from rumors, and weak uncertainty calibration. It also warned of possible formatting and citation errors. A report with many citations can still misread a source or overstate what the evidence supports.
For consequential work, check key claims against the linked original material, verify dates and definitions, and examine whether cited evidence actually supports the conclusion. Distinguish a documented fact from the model’s inference. This is particularly important when research informs medical, legal, financial, or other high-stakes decisions.
Access changes after February 2025
- February 2, 2025: the initial rollout targeted ChatGPT Pro with up to 100 queries per month, with other plans announced to follow.
- February 5, 2025: OpenAI expanded Pro access to the United Kingdom, Switzerland, and the European Economic Area. The original version of this article incorrectly continued to describe those regions as unavailable.
- February 25, 2025: OpenAI announced access for all Plus users.
- April 24, 2025: a lightweight o4-mini-based version expanded access and changed the published allowances. Those April allowances are also historical rather than a promise of current quotas.
Mobile and desktop support, richer reports, and connected data sources were described in the launch roadmap. They should not remain indefinitely labeled as features arriving within the next month.
Current ChatGPT feature versus retired API models
OpenAI’s current ChatGPT guide, checked October 19, 2026, describes a reviewable research plan, progress controls, public-web and uploaded-file research, and eligible connected-app sources. App availability depends on plan, region, workspace configuration, and permissions; an available app does not automatically support deep research. The guide says research uses app read actions.
Usage varies by plan. The in-product counter shows remaining tasks; fixed monthly allowances reset every 30 days from first use. The guide says the latest model is used by default. Use the account’s current interface and help documentation rather than the old Pro-only limit or the launch model name to determine today’s access.
For API developers, the API deep research guide explicitly labels its old model examples as reference material. The deprecation record gives July 23, 2026 as the shutdown date for the June 26, 2025 snapshots and aliases of o3-deep-research and o4-mini-deep-research, listing gpt-5.6-sol as the replacement. This model-level change is separate from continued deep research in ChatGPT.
The launch’s historical importance was its shift toward longer, source-backed AI investigation. The practical standard remains the quality and traceability of the evidence, together with appropriate human review.
