OpenAI is launching a brand new common function AI agent in ChatGPT, which the corporate says can full all kinds of computer-based duties on behalf of customers. OpenAI says the agent can robotically navigate a consumer’s calendar, generate editable displays and slideshows, and run code.
The software, referred to as ChatGPT agent, combines a number of capabilities from OpenAI’s earlier agentic instruments, together with Operator’s ability to click around on websites, in addition to Deep Research’s ability to synthesize information from dozens of internet sites right into a concise analysis report. OpenAI says customers will be capable of work together with the agent just by prompting ChatGPT in pure language.
ChatGPT agent is rolling out on Thursday to subscribers to OpenAI’s Professional, Plus, and Staff plans. To activate the software, customers can choose “agent mode” in ChatGPT’s dropdown menu of instruments.
The launch of ChatGPT agent represents OpenAI’s boldest try but to show ChatGPT into an agentic product that may take actions and offload duties for customers, quite than simply answering questions. Lately, Silicon Valley firms together with OpenAI, Google, and Perplexity have unveiled dozens of AI brokers which have promised to just do that. Nevertheless, these early version AI agents have proven to struggle with complex tasks, they usually appear much less compelling as merchandise than the final word imaginative and prescient tech executives pitch round AI brokers.
That mentioned, OpenAI says ChatGPT agent is much extra succesful than its earlier choices.
The corporate’s new agent can entry ChatGPT connectors, permitting customers to attach apps like Gmail and GitHub in order that the agent can discover related info to your prompts. OpenAI says ChatGPT agent has entry to a terminal, and it will probably use APIs to entry sure apps.
OpenAI means that customers can faucet ChatGPT agent to “plan and purchase elements to make Japanese breakfast for 4,” in addition to “analyze three opponents and create a slide deck.” These sorts of capabilities requires ChatGPT agent to parse via web sites, plan a plan of action, and use instruments — rather more difficult duties than OpenAI has beforehand tried to deal with with brokers.
Techcrunch occasion
San Francisco
|
October 27-29, 2025
The mannequin underlying ChatGPT agent provides state-of-the-art efficiency on a number of benchmarks, in keeping with OpenAI.
The corporate says the ChatGPT agent mannequin scores 41.6% on Humanity’s Final Examination (go@1), a troublesome take a look at made up of hundreds of questions throughout a couple of hundred topics. That’s roughly double what OpenAI’s o3 and o4-mini scored on the take a look at.
On FrontierMath, one of many hardest identified math benchmarks, OpenAI says ChatGPT agent scores 27.4% when it has entry to instruments, similar to a terminal for code execution. The earlier state-of-the-art rating comes from o4-mini, which scored simply 6.3%.
OpenAI notes that it developed ChatGPT agent with security in thoughts, largely as a result of the product presents some newfound capabilities that would make it extra harmful within the fingers of a foul actor. OpenAI has beforehand warned that agentic fashions might current extra harmful capabilities.
In a safety report for ChatGPT agent, OpenAI says it’s designated the mannequin as “excessive functionality” in organic and chemical weapon domains, which is outlined in OpenAI’s Preparedness Framework as a mannequin with the power to “amplify current pathways to extreme hurt.” OpenAI notes that it doesn’t have direct proof of this, but it surely’s determined to take a precautionary method and activate new safeguards to mitigate these dangers.
The brand new safeguards for ChatGPT agent embrace a monitor that works in actual time as customers work together with the product. OpenAI says it runs a classifier throughout each immediate entered into ChatGPT agent, figuring out whether or not the request is said to biology. If that’s the case, OpenAI runs ChatGPT agent’s response via a second monitor that determines whether or not the content material might be used to evoke a organic menace.
OpenAI additionally says it disabled ChatGPT’s reminiscence characteristic for this agent to forestall misuse. In different components of ChatGPT, OpenAI’s reminiscence characteristic permits the chatbot to reference info from earlier consumer chats. Nevertheless, OpenAI says dangerous actors might use the characteristic in ChatGPT agent to exfiltrate delicate information via immediate injection assaults. The corporate says it could revisit including the characteristic sooner or later, nevertheless.
Whereas ChatGPT agent sounds spectacular, it stays to be seen how succesful it really is in the actual world. Till now, agent expertise has confirmed comparatively brittle when interacting with the actual world. That mentioned, OpenAI says it’s developed a extra succesful mannequin that’s capable of ship on the promise of AI brokers.
This story was up to date with extra info.