OpenAI launches a general purpose agent in ChatGPT

OpenAI is launching a new AI agent for general purposes in chatgpt, the company says it can complete a wide range of computer-based tasks on behalf of users. OpenAI says that the agent can automatically navigate the agenda of a user, generate editable presentations and slide performances and perform code.

The Tool, called Chatgpt -Agent, combines various options from the earlier agent tools of OpenAI, including the capacity of the operator to click around on websites, as well as the ability of deep research to synthesize information from dozens of websites in a concise research report. OpenAi says that users will be able to communicate with the agent by simply requesting chatgpt in natural language.

Chatgpt agent will be rolled out on Thursday for subscribers to OpenAi’s Pro, Plus, and team plans. To activate the tool, users can select the “Agent mode” in the Chatgpt with tools drop -down menu.

The launch of the chatgpt agent represents the bravest attempt by OpenAi to turn Chatgpt into an agentic product that can take actions and can discharge tasks for users, instead of just answering questions. In recent years, companies in Silicon Valley, including OpenAi, Google and Parmexity, have unveiled dozens of AI agents who have promised to do that exactly. However, this early version AI agents have proven that they are struggling with complex tasks, and they seem less attractive as products than the ultimate executives of Vision Tech around AI agents.

That said, OpenAi says that Chatgpt agent is much more capable than his earlier offers.

The new agent of the company has access to chatgpt -connectors, so that users can connect apps such as Gmail and Github, so that the agent can find relevant information for your prompts. OpenAi says that Chatgpt agent has access to a terminal, and it can use APIs to gain access to certain apps.

OpenAI suggests that users can tap a chatgpt agent to “plan and buy ingredients to make Japanese breakfast for four”, as well as “analyze three competitors and make a slide.” These types of possibilities require that ChatGPT agent parses via websites, plan a course of action and use tools – much more complicated tasks than OpenAi has tried to tackle agents before.

The underlying model-underneath ChatGPT agent offers state-of-the-art performance on various benchmarks, according to OpenAi.

The company says that the Chatgpt -agent model scores 41.6% on the last exam of humanity (Pass@1), a difficult test consisting of thousands of questions about more than a hundred topics. That is about double what OpenAi’s O3 and O4-Mini scored on the test.

On Frontiermath, one of the most difficult mathematical benchmarks, OpenAI says that Chatgpt Agent scores 27.4% when it has access to tools, such as a terminal for code version. The previous state-of-the-art score comes from O4-Mini, who scored only 6.3%.

OpenAI notes that the Chatgpt agent has developed with safety in mind, largely because the product presents some new possibilities that can make it more dangerous in the hands of a bad actor. OpenAi has previously warned that agent models could offer more dangerous options.

In one Safety report For Chatgpt -agent, OpenAI says it has designated the model as “high capacity” in biological and chemical weapons domains, which is defined in the openness framework of OpenAi as a model with the ability to “strengthen” existing paths to serious damage “. OpenAI notes that it has no direct evidence of this, but it has been decided to follow precautions and activate new guarantees to reduce these risks.

The new guarantees for Chatgpt agent include a monitor that works in real time while users deal with the product. OpenAI says it is performing a classifier in any promptly introduced in Chatgpt agent and determines whether the request is related to biology. If this is the case, OpenAI carries out Chatgpt Agent’s reaction through a second monitor that determines whether the content can be used to evoke a biological threat.

OpenAi also says that it has eliminated the memory function of chatgpt for this agent to prevent abuse. In other parts of chatgpt, the memory function of OpenAI can refer the chatbot to information from earlier user treasure. However, OpenAi says that bad actors can use the function in chatgpt agent to exfil sensitive data due to fast injection attacks. However, the company says that it can be visited again when adding the position in the future.

Although Chatgpt -agent sounds impressive, it is still to be seen how capable it really is in the real world. Until now, Agent Technology has proved to be relatively brittle when dealing with the real world. That said, OpenAi says it has developed a more capable model that can make the promise of AI agents come true.

This story has been updated with more information.