July 18th.OpenAI It was announced early this morning that the ChatGPT The launch of aGeneral Purpose AI AgentThe company says the smart body can help users with a variety of computer-based tasks.
OpenAI describes the intelligence as being able to automatically generate editable presentations and slideshows, view a user's calendar to give a brief overview of an upcoming client meeting, plan and purchase ingredients to make a family breakfast, and run code, among other things.
The tool is called ChatGPT agent.Combines the functionality of many of OpenAI's previous intelligences toolsOpenAI says users can interact with the intelligence simply by using the natural language prompt ChatGPT, which includes the ability for Operator to click on a website and for Deep Research to synthesize information from dozens of websites to produce a succinct research report.
In order to develop the new tool, OpenAI has merged the Operator and Deep Research teams behind it into one unified team. The Verge reports that the new team consists of 20-35 people from the product and research departments.
OpenAI says the ChatGPT Intelligence is much more powerful than any of its previous products, and can access the ChatGPT Connector.Allows users to connect to apps like Gmail and GitHubChatGPT intelligences can find relevant information based on user prompts. In addition, OpenAI says the ChatGPT intelligences have access to terminals and can use APIs to access certain applications.
According to OpenAI, the underlying model of the ChatGPT intelligences provided state-of-the-art performance across multiple benchmarks.The ChatGPT intelligences model scored 41.6% on Humanity's Last Exam (pass@1), a difficult test comprised of thousands of questions covering more than one hundred subjects on the difficult test.This score is roughly double the OpenAI o3 and o4-mini scores!.
In FrontierMath, one of the hardest known math benchmarks, OpenAI says that the ChatGPT intelligence scored 27.41 TP3T when it had access to tools (note: such as terminals used for code execution), with the previous best score coming from o4-mini (which scored just 6.31 TP3T).
In the DSBench test, which evaluates the performance of intelligences in realistic data science tasks covering data analysis and modeling, the ChatGPT intelligences significantly outperformed previous state-of-the-art models - especially in data analysis tasks, where they significantly outperformed the human level. The ChatGPT Intelligence
On the SpreadsheetBench platform, which scores models by evaluating their performance when handling spreadsheet editing tasks based on real-world scenarios, the ChatGPT Intelligence set a new industry-leading level (SOTA), more than doubling the performance of the current industry-leading GPT-4o. When given the ability to edit spreadsheets directly, the ChatGPT Intelligence further improves its score to 45.5%, matching Copilot's 20.0% in Excel.
In an internal benchmark test, the model demonstrated its ability to handle tasks for investment banking analysts (1 to 3 years of experience), such as modeling specification-compliant financial statements (including formatting and citation) for a Fortune 500 company or modeling a leveraged buyout for a going-private transaction. the model used by the ChatGPT intelligences significantly outperformed the deep-dive and the o3 models in this test. Each task is scored on hundreds of criteria related to correctness and formula usage.
In the WebArena benchmark test, which evaluates the performance of web browsing intelligences in accomplishing real-world web tasks, this model performs better compared to the o3-driven CUA, the model that drives Operator.
ChatGPT intelligences on BrowseComp (a benchmark test released by OpenAI earlier this year) to measure the ability of browsing intelligences to find hard-to-find information on the web.The model set a new SOTA record with 68.9%which is 17.4% higher than Deep research.
Specific use scenarios:
- At work, users can automate repetitive tasks such as converting screenshots or panels into presentations consisting of editable vector elements, rescheduling meetings, planning and booking outings, and updating spreadsheets with new financial data while maintaining the original formatting.
- In their personal lives, users can plan and book a travel itinerary, design and book an entire dinner event, or find a professional and schedule an appointment.
In terms of security, OpenAI says users will always be in control.ChatGPT will ask the user's permission before performing important operationsThe user can interrupt the operation, take over the browser or stop the task at any time.
Users can activate ChatGPT's new Intelligentsia feature directly from the Tools drop-down menu in the editor by simply selecting "intelligent body model" is sufficient. Simply describe the task you wish to accomplish - whether it's conducting in-depth research, creating a slide deck, or submitting an expense claim. As the task is performed, an on-screen voice announcement shows the exact flow of ChatGPT in real time. Users can interrupt and take over the browser operation at any time, ensuring that the task remains on target.
In addition, users can set completed tasks to repeat automatically, such as automatically generating weekly metrics reports every Monday morning.
ChatGPT Intelligent Body is available now to Open for Pro, Plus and Team Edition usersThe Enterprise and Education editions will be available in July, with nearly unlimited tasks per month for Pro Edition users and 50 tasks per month for other paid users, with additional usage available through flexible credit options.
OpenAI stated thatChatGPT Intelligent Bodies Still in Early Stages -- It is capable of handling a wide range of complex tasks, but errors can still occur. Although the feature is officially recognized as having great potential for generating slideshows, it is still in beta -- currently generated content may appear rough in terms of formatting and detailing, especially when starting to create without an existing document. Additionally, while you can currently upload existing spreadsheets for ChatGPT to edit or use as templates, this feature is not yet available for slideshows.
OpenAI is training a next-generation version of the ChatGPT Slide creation feature to generate more polished and complex output with a wider range of features and improved formatting capabilities.
OpenAI plans to add major improvements incrementally at a regular paceand make ChatGPT intelligences more and more useful to more people over time.