OpenAI is diving deeper into AI agents with its latest offering.
In a livestreamed announcement from Tokyo, the AI powerhouse introduced Deep Research, a new capability that lets AI models conduct extensive internet research and synthesize information—without the usual time constraints that limit most AI systems.
So what makes it different?
While typical AI models return responses quickly, Deep Research can spend up to 30 minutes browsing the web, analyzing content, and compiling comprehensive reports.
The system is powered by OpenAI’s upcoming o3 reasoning model and was trained using end-to-end reinforcement learning.
Deep Research is flexing some impressive muscles on industry benchmarks:
- Achieved 26.6% accuracy on Humanity’s Last Exam, setting a new high score
- Outperformed previous models across all difficulty levels on the GUIA benchmark
- Demonstrated expert-level performance on internal evaluations, completing tasks that would take human experts hours
The feature is rolling out today to ChatGPT Pro subscribers, with plans to expand to Plus and Team users soon. Education and Enterprise access will follow.
This release signals OpenAI’s broader vision for AI agents that can work autonomously for extended periods.
The company sees Deep Research as a stepping stone toward systems that can uncover and generate new knowledge independently—though they recommend users verify sources, as hallucinations are still possible.
