OpenAI ships Deep Research

OpenAI launched Deep Research, a ChatGPT feature that took a research question, browsed and read sources autonomously for five to thirty minutes, and returned a written report with citations. It was built on a version of the o3 reasoning model fine-tuned specifically for browsing and synthesis, and could read text, images and PDFs encountered along the way. Access was initially limited to ChatGPT Pro subscribers, at $200 a month, capped at 100 queries a month; broader tiers followed later.

OpenAI reported that the model powering Deep Research scored 26.6% on Humanity’s Last Exam, a benchmark of expert-level questions across hundreds of subjects then considered one of the hardest available, well above GPT-4o’s 3.3% and DeepSeek’s R1, which had scored around 9.4% weeks earlier. The company and outside coverage framed the jump as evidence that giving a model time and tools to research, rather than answering from a single pass, produced large capability gains on knowledge-intensive tasks.

OpenAI also acknowledged clear limitations: the system could still “hallucinate facts or make incorrect inferences,” sometimes presenting unverified claims without adequately signalling uncertainty, and its long, dense outputs made errors harder for a user to catch than a short answer would be. Google had shipped a broadly similar “Deep Research” feature in Gemini two months earlier, in December 2024, without the same benchmark claims attached, so OpenAI’s release consolidated rather than originated the category.

The launch is generally treated as the point at which long-horizon, tool-using research agents moved from lab demo to a product people actually used, prompting rapid competition: Perplexity, Google and others expanded their own research-agent offerings within the following months, and “deep research” became a standard feature name across chatbot products.