To develop Deep Research, we created reinforcement learning tasks to teach the model browsing and data analysis capabilities, and provided it access to tools for training.

The model that powers Deep Research is a fine-tuned version of the O3 model, specifically trained on challenging browsing and reasoning tasks, enhancing its online search and analysis capabilities.

Deep Research is an agentic capability in ChatGPT that conducts multi-step online research to solve complex tasks, delivering a fully cited comprehensive report in minutes, which would take a human many hours.

Research indicated that previous models like GPT-3 were undertrained; despite their size, they hadn't been trained on enough text to reach their full potential. For instance, Chinchilla, a model smaller than GPT-3, outperformed larger models by being trained on four times more data.

We aimed to create a useful model by training it directly on tasks that users perform in their daily lives, with online browsing being a key focus due to its broad applicability.

Deep Research's ability to train end-to-end allows it to adapt flexibly to live web information, making it capable of creative searches and generating insightful responses.

We started building Deep Research over a year ago, driven by progress in reinforcement learning and reasoning models, initially focusing on math, science, and coding tasks.

GPT-3 made headlines by smashing benchmarks previously considered out of reach for AI, excelling in software engineering, math, and PhD-level science questions.

OpenAI's release of GPT-2 in November 2019 marked a significant milestone with 1.5 billion parameters, followed by GPT-3, which was over 100 times larger and more capable than its predecessor.