Chatgpt’s Agent Can Now Do Deep Research For You

OpenAI has revealed different caller agentic characteristic for ChatGPT called dense research, which it says tin tally autonomously to “plan and execute a multi-step trajectory to find nan accusation it needs, backtracking and reacting to real-time accusation wherever necessary.”

Instead of simply generating text, it shows a summary of its process successful a sidebar, pinch citations and a summary showing nan process utilized for reference.

Users tin inquire questions utilizing text, images, and further files for illustration PDFs aliases spreadsheets to adhd context, and past it will return “anywhere from 5 to 30 minutes” to create a consequence provided successful nan chat window, pinch promises that successful nan early it will too beryllium tin to spot embedded images and charts. OpenAI too notes limitations for dense research, saying it tin “sometimes hallucinate” and dress up facts, struggle pinch telling nan value betwixt charismatic info and rumors, and registry really definite it should title a response.

Developing ways for generative AI devices to beryllium overmuch useful and worthy paying for is nan early companies for illustration OpenAI person promised for agents, and it claims that dense investigation is tin of operating astatine nan level of a investigation analyst. The demo video included coming originates pinch a petition for info connected changes successful nan portion manufacture complete nan past 3 years, pinch a consequence that includes slug points and tables.

This characteristic intimately follows OpenAI’s motorboat of Operator, a instrumentality that tin usage a web browser to complete tasks for you, and is akin to nan Project Mariner investigation prototype Google showed disconnected successful December. Google’s instrumentality is not disposable to nan nationalist yet, but dense investigation is launching “with a type optimized for Pro users today.”

OpenAI is offering up to 100 queries per play for those paying nan $200 monthly liking and “limited access” promised for Plus, Team, and eventually, Enterprise users, calling nan expertise “very compute intensive,” requiring overmuch conclusion compute nan longer it takes to investigation something. It too says that each paid users will get higher title limits successful nan early erstwhile a faster, overmuch cost-effective type is available.

A spot merchandise says that nan exemplary powering dense investigation scored a caller precocious for accuracy connected an AI benchmark dubbed “Humanity’s Last Exam,” which asks for responses to expert-level questions. The OpenAI dense investigation exemplary reached an accuracy of 26.6 percent pinch browsing and python devices enabled, bully supra GPT-4o’s 3.3 percent, and nan adjacent highest scorer, its o3-mini (high) exemplary evaluated only connected text, astatine 13 percent.