OpenAI Launches Research Analysis AI Agent

OpenAI has launched an artificial intelligence (AI) agent it says can emulate the work of a research analyst for areas including finance, science, policy and engineering, carrying out in “tens of minutes” what would take a human “many hours”.

The Deep Research capability is offered through the firm’s ChatGPT chatbot and uses a version of the upcoming o3 “reasoning” model optimised for web browsing and data analysis.

It uses the model’s reasoning capabilities to search, interpret and analyse massive amounts of text, images and PDFs from the internet and produce a report within the chat.

The report will later include embedded images, data visualisations and other analytic output, OpenAI said.

‘Very compute intensive’

Deep Research is the second of OpenAI’s “agent” capabilities, following Operator, designed to help book flights, plan grocery orders based on an image of a shopping list, and even complete purchases with minimal supervision.

Both services are initially available to users of the recently introduced ChatGPT Pro plan who pay $200 (£161) per month for access to cutting-edge features.

Users are to be limited to 100 queries a month, reflecting the high cost of carrying out multi-step tasks using the o3 model, which itself consumes significantly more processing power than previous versions.

The research tool is “very compute intensive”, OpenAI said.

OpenAI plans to eventually offer the agent capabilities to other plans such as Plus, Team and Enterprise, but did not provide a timeline for doing so.

The firm said the agent was a “significant step” toward the goal of artificial general intelligence, the concept of AI that can perform general tasks at least as well as a human.

But it cautioned that Deep Research is still in its early stages and can present fabricated information as factual and may not be able to distinguish rumours from accurate information – common features of generative AI tools.

Productivity boost

Such shortcomings are an indicator of the limitations of AI’s promise to make human workers more productive, since their productions necessarily require time-intensive human oversight to ensure their accuracy.

Operator, introduced in December, is similarly available only as a preview version. Both are currently offered only in the US.

A demonstration video released by OpenAI showed Deep Research producing a report on the market for translation apps. A report could take from five to 30 minutes to produce a report and will cite a source for each claim it makes, the company said.

In addition to professional uses, the tool could create comparison reports for individuals looking to purchase cars or furniture.

On Friday OpenAI released another o3 derivative, a free slimmed-down version called o3-mini.

OpenAI on Monday announced a joint venture with Japan’s SoftBank aimed at marketing AI services to Japanese corporate customers.