ChatGPT Agent: Not surpassing Manus' capabilities, but seeing the dawn of end-to-end

36 duty assistant

·July 19, 2025 09:34

Operator and Deep Research merge

Agent is the biggest consensus in the AI community this year, and OpenAI naturally cannot fall behind.

At 1am on July 18, 2025 Beijing time, Sam Altman and four OpenAI researchers officially released ChatGPT Agent - a general-purpose AIAgent - during a live broadcast.

Previously, there were Manus, Lovart, and Flowith. The functional scenarios presented by ChatGPT Agent were not particularly stunning, but the significance of its release goes beyond its functionality itself.

The revolution of ChatGPT Agent lies in its unique technological path: it can actively select agent skills from the toolbox, complete tasks using its own computer, and users can observe the AI's working process in a virtual environment in real time.

Although this interactive interface is similar to products such as Manus, the underlying principles have fundamental differences. Manus calls multiple underlying models, similar to "external stitching", while ChatGPT Agent internalizes agent capabilities into the model. We have seen the prototype of an end-to-end universal agent.

Manus' design essentially achieves' external stitching 'by calling multiple underlying models. In contrast, ChatGPT Agent internalizes its agent capabilities within the model itself.

According to OpenAI, in order to develop the ChatGPT Agent, they merged the Operator and Deep Research teams into a unified team consisting of 20 to 35 people.

According to the system card of ChatGPT Agent, it is a new agent model belonging to the same series as OpenAI o3, and adopts an end-to-end training method. It is a unified model developed for agent tasks, rather than an engineering combination of multiple models.

According to the comparison PPT released by OpenAI, we can see that this training is basically completed through the reinforcement learning process. The path should be similar to Grok4withtool.

After retraining, the Agent combines the multi-step research and high-quality report generation capabilities of Deep Research, the ability of Operators to execute tasks through remote visual browser environments, terminal tools with limited network access, and the ability to access external data sources and applications through connectors.

After completing complex tasks, a downloadable PPT or document can also be delivered to the user.

For Manus, OpenAI's new initiative is undoubtedly a huge blow, and even in terms of pricing, the difference between the two is not significant: the GPT Plus package can use ChatGPT Agent for $20 per month, while Manus' basic plan is $19 per month.

Key points:

ChatGPT Agent: A unified AI agent capable of executing complex, multi tool tasks.

It integrates access to text browsers, GUI browsers, terminals, and image generation tools.

Support interactive and multi round conversations with users, allowing interruptions and clarifications.

Security Protection Upgrade: Strengthen the defense against "malicious prompts" attacks on web pages; Set up automatic rejection of high-risk tasks; Biological/chemical risks are also handled according to the highest level of safety stack.

It has achieved state-of-the-art results in multiple real-world and benchmark tasks.

ChatGPT Agent Overview: Functionality Similar to Manus

The core of ChatGPT Agent is a unified agent system that integrates and extends the capabilities of OpenAI's early research projects "Operator" (focusing on website interaction) and "Deep Research" (focusing on information synthesis).

This enables ChatGPT Agent to seamlessly switch from reasoning and thinking to executing specific actions within a single conversation flow.

Virtual computer environment: ChatGPT Agent executes all tasks on a specially designed virtual computer. This environment is sandboxed to ensure the security of operations. It can save the context of tasks in the environment, and even if the user interrupts or changes instructions midway, it can continue from the breakpoint without losing progress.

Intelligent Toolbox: In order to complete complex workflows, the Agent is equipped with four tools and can automatically select the most suitable tool based on task requirements:

Visual Browser: Used for interacting with graphical user interfaces, such as clicking buttons, filling out forms, and browsing websites designed for humans.

Text based Browser: Used for network queries that require efficient inference and processing of large amounts of text.

Terminal: Allow agents to run code, download and process files.

API access: You can directly call APIs to obtain information, such as accessing data from applications like Google Drive, Gmail, and GitHub through connectors.

New Model Driven: ChatGPT Agent is driven by a new model specifically developed for it. This model has been specifically trained on complex tasks that require the use of multiple tools through reinforcement learning methods, thus learning how to smoothly switch and collaborate between different tools.

It has the following characteristics:

Autonomous task execution: Users can issue instructions in natural language, such as "analyze my calendar and brief me on upcoming client meetings based on recent news." Agents can autonomously plan and execute a series of operations, such as browsing websites, filtering information, running code analysis, and ultimately generating editable slides or spreadsheets.

Collaboration and Interactivity: It will proactively ask for more details when needed to achieve goals. Users can interrupt, redirect tasks, or take full control of the browser at any time.

Security and permission control: Security is the core part of its design. Before performing critical operations that have practical impact, such as purchasing, submitting forms, sending emails, or processing personal information, the agent will explicitly request user permission. Meanwhile, it is prohibited from carrying out high-risk tasks such as financial transfers or providing legal advice. OpenAI also has built-in protection measures against malicious attacks such as "prompt injection".

Multiple benchmark test scores break records

The most difficult HLE reached 41.6% (with tool), higher than the just released Grok4 (with tool) at 41.0%.

On Humanity's Last Exam, which measures wide area knowledge and expert level questioning, the accuracy rate of a single answer reached 41.6%; By using parallel eight way reasoning and selecting the answer with the highest confidence level, it can be improved to 44.4%.

On the extremely difficult FrontierPath mathematical benchmark, the accuracy was improved to 27.4% after running the code with the terminal.

In the internal evaluation of real knowledge work tasks, ChatGPT agents have been on par or better than humans in about half of the cases;

On the real-world data science task DSBench, its analysis and modeling accuracy reached 89.9% and 85.5%, respectively, far exceeding the human average level.

It also leads in direct editing ability for spreadsheets: achieving 45.5% in Spreadsheet Bench, surpassing Copilot in Excel's 20%. In addition, it has refreshed SOTA in browsing reviews such as BrowseComp and WebArena.

(Image: Evaluation method: The author of Spreadsheet Bench evaluated the spreadsheet using Microsoft Excel in a Windows environment.). We use LibreOffice in the OSX environment, which may result in slight differences in ratings. For example, the author reported a result of 15.02% for GPT-4o on the overall Hard limit, while we obtained 13.38%. We used a complete benchmark test of 912 questions. ）

According to the PPT created by ChatGPT Agent, the agent's ability to create PPT and surf the internet has significantly improved compared to the pure basic model. But there is still a long way to go from humans.

Not futures, available today

Starting today, Pro users can use it immediately, while Plus and Team users will gradually activate it within a few days; The Enterprise and Education versions will be integrated in a few weeks.

Pro can receive 400 messages per month, while other paying users have a monthly limit of 40 messages, which can be added through flexible pay as you go billing.

The actual use is very simple: switch to "proxy mode" in any conversation, describe the goal, such as conducting in-depth research, making presentations, or reimbursement. Real time display of its operation process on the left side of the screen; If login is required, the system will switch to "takeover mode" for secure input of credentials.

Users can also set the completed tasks to be executed periodically, such as automatically generating indicator reports every Monday.

Ultraman personally warns of risks: Agents are powerful and dangerous

It is worth noting that Ultraman immediately posted a long post after the press conference, warning about the risks of using ChatGPT Agent.

After emphasizing the powerful ability of ChatGPT Agent to handle complex tasks, we solemnly pointed out the risks of the product and emphasized that we are not yet clear about the specific impact, but criminals may attempt to "trick" users' AI agents into providing private information that should not be provided and taking actions that should not be taken, and we cannot predict the ways in which this will happen.

The model may come into contact with sensitive user data or encounter malicious' prompt injection 'attacks in web pages. For this reason, they continue to use the strict control during the Operator period and have added multiple protections:

Clear user authorization must be obtained before key actions are taken;

Enabling 'supervision mode' for some high-risk tasks (such as sending emails) requires users to monitor the entire process;

Will actively refuse high-risk instructions such as bank transfers;

Users can easily clear browsing data and log out of all sessions with just one click, or disable the connector when not connected to the internet.

In terms of biological and chemical security, OpenAI has classified the model as high-risk based on the Preparedness Framework, implemented the most comprehensive security measures, and collaborated with government, academia, and security agencies to conduct red team testing and threat modeling. At the same time, OpenAI has launched a vulnerability bounty program to detect and patch potential issues as early as possible.

Is ChatGPT Agent far ahead enough?

The biggest innovation of ChatGPT Agent is the direct integration of a complete virtual machine environment into the model for the first time, allowing users to observe the AI's operation process in real time, which is not available in other model products.

However, mainstream model companies are moving further and further down the path of 'Agent as Model, Model as Agent'. For example, Claude, who is almost a god in coding agent abilities.

Many Agent products that require borrowing underlying models to build, even without Claude, are nothing.

The newly launched Kimi K2 adopts an open-source hybrid expert model architecture, positioned as Agentic Intelligence, and priced at only about 1/6 of Claude 4. After going online, the adoption ranking of tokens continued to soar.

But in terms of the path of 'model as agent', OpenAI cannot be considered far ahead, it can only be said to have taken a small step forward.

OpenAI also humbly stated in its official documentation that:

It should be noted that the functionality is still in its early stages: for example, the slide generation function is currently in beta, and the format and aesthetics still need to be improved. At this stage, the main focus is on optimizing the information structure and element editability; In the future, we will continue to train new versions to generate more refined files. Overall, with continuous iteration, the efficiency, depth, and diversity of ChatGPT agents will continue to improve, and we will gradually optimize the strength of user supervision to achieve a better balance between usability and security.

Watching the demonstration of his own product, Sam Altman couldn't help but exclaim, 'I feel AGI.'.

However, there were still user comments after the post asking, 'What about the GPT-5 that was agreed upon?'?

This article is from the WeChat official account "Tencent Technology", written by Xiao Jing and Bo Yang, 36 Krypton has been authorized to release.

The viewpoint of this article only represents the author himself, and the 36Kr platform only provides information storage space services.

The images in this article are taken by the author, provided through interviews, and authorized by the company

ChatGPT Agent：没超越Manus的能力范畴，但看到了端到端的曙光