OpenAI dropped GPT-5.4 without warning.This model will change the entire world in 2026AIAgent pattern!This time, the core ability is finally ready to be revealed:Native computer control.Agent capabilities are the main task of AI evolution in 2026. After the actual test, I have a profound feeling:GPT-5.4 may be the most suitable model for running OpenClaw so far in 2026.Maybe not one.
First-hand measured GPT-5.4
The workers are really in danger
At present, GPT-5.4 has been launched in the web version and Codex. After testing various scenarios, there is really only one feeling:The workers are really in danger.
This model can operate almost everything on the computer: files, programs, files, terminals, etc., as long as a jobber can do it, it can do it.
Taking a Macbook as an example, let GPT-5.4 operate the calendar to customize reminders.
GPT-5.4 directly launches the corresponding application and requires the system to grant corresponding permissions.
Then the corresponding reminder appears in the calendar.
You can also use GPT-5.4 to operate the computer to open the APP, such as asking to open the Little Universe APP.
GPT-5.4 can not only find APPs, but also play corresponding programs.
You can even ask GPT-5.4 to directly operate the calculator APP, and it isInside the calculator APPto calculate.
Almost any APP on your computer can be called by GPT-5.4, such as opening WeChat to read.
This is true "native computer control capability."
You can ask GPT-5.4 to change your computer wallpaper.
Even more, you can ask GPT-5.4 to operate the terminal to open the claude code.
We all know that the "lobster" OpenClaw can become popular because its Agent capabilities are too powerful.
An OpenClaw deployed on a Macmini has almost the same operating capabilities and operating permissions as a human.
This time, GPT-5.4 directlyat model levelRealize the ability to "natively control the computer".
Peter, the father of OpenClaw, commented that GPT-5.4 has greatly improved its programming capabilities and is more unified and smarter in other aspects.
Matt Shumer, CEO of HyperWriteAI and OthersideAI, believes that GPT-5.4 is currently the best model in the world, far surpassing others.
His evaluation of GPT-5.4 is:Incredible, beyond imagination.
Even in standard mode, GPT-5.4's performance surpasses previous professional models... This is incredible!
His programming skills are even more amazing...almost flawless.
Internally, Codex is as reliable as it gets.
The programming problem has largely been overcome. Needless to say about this, it is so outstanding that it is beyond imagination.
Sebastien Bubeck, a distinguished scientist at OpenAI, used a meme to describe the improvement in GPT-5.4 programming capabilities.
Brenda, CEO of Mercor AI, believes that GPT-5.4 will soon surpass the professional capabilities of top consulting firms, top investment banks, and top law firms.
That is to say, it can do all the work of the human elite golden collar, and even do it better.
Native computer control capabilities
This is a real watershed
After the actual measurement, let us think carefully, what is the strength of GPT-5.4?
GPT-5.4 is the first general-purpose model released by OpenAI with "native computer usability".
It can directly identify screen content through screenshots, issue mouse clicks and keyboard input commands, and operate software in a desktop environment like humans without resorting to external frameworks or specialized adaptation layers.
In the OSWorld-Verified benchmark test, which measures the AI's ability to navigate a real desktop environment, GPT-5.4 achieved a success rate of 75.0%.
For comparison, GPT-5.2 is only 47.3%, while the human baseline is 72.4%.
In other words, GPT-5.4's ability to control computers has exceeded the average level of ordinary human users.
This is the first time a general-purpose AI model has done this.
AI controls computers, for the first time, better than humans themselves.
A match made in heaven: GPT-5.4 accurately hits every pain point of OpenClaw
Back to OpenClaw, why is GPT-5.4 said to be OpenClaw’s chosen model?
Finally came the most exciting part.
If you paid attention to the technology circle at the beginning of this year, it is impossible for you not to know about OpenClaw.
This open source personal AI assistant project developed by Peter Steinberger, the founder of PSPDFKit, became popular overnight in January 2026 at a rate of 60,000 GitHub stars in 72 hours. Now the number of stars has exceeded 250,000.It even once caused Mac Mini to be sold out globally and the Raspberry Pi stock price to double.
The core philosophy of OpenClaw is just one sentence:TheAIthat actually does things.
Not just chatting, but actually helping you work.
It runs on your own computer and has full access to the shell, file system, and browser.
You send a message through WhatsApp, Telegram, Feishu, and DingTalk, and it can help you execute terminal commands, write scripts, manage emails, capture web page data, and even install and debug software in the background.
It has a lasting memory and can become more and more powerful by continuously expanding its skills.
However, OpenClaw has an open secret:The model is the product.
OpenClaw founder Peter Steinberger himself said,It is recommended to use the strongest latest generation model.
The community’s conclusion is also sharp:The model is the product, and the framework is just the packaging.
A widely circulated experience in the community is thatIf the model you use is not strong enough, the magic of OpenClaw will disappear.
Some developers have done detailed tests, and the experience gap between running OpenClaw on a low-end model and Claude Opus 4.5 can be 40% to 95%.
Peter, the founder of OpenClaw, recommends upgrading the OpenClaw version to 5.4 and then working with GPT-5.4.
The reasoning ability of the model, the accuracy of tool calling, and the ability to maintain the context of long-range tasks directly determine whether OpenClaw is a cyber secretary or an expensive automatic completion.
GPT-5.4 hits the sweet spot in every dimension that OpenClaw needs most.
Native computer control - a perfect match.
The core value of OpenClaw is to let AI operate your computer. GPT-5.4 is the first general-purpose model with this capability natively, and OSWorld tests exceed human levels. This means that OpenClaw, which is connected to GPT-5.4, no longer needs to implement desktop automation through complex adaptation and Hack, but is like replacing a turbine engine with a racing car - it was born for this scenario from the bottom up.
1 million Token context - worry-free battery life.
OpenClaw is a continuously running Agent that needs to maintain long conversations, remember complex command chains, and process large amounts of file content. In the past, when using other models, I would forget the previous tasks while running. The context window of 1 million Tokens means that OpenClaw finally has a workbench large enough to spread out all the materials.
Tool Search——Agent’s efficiency revolution.
The power of OpenClaw lies in its access to dozens of skills and tools. But the more tools there are, the more Tokens are consumed for each call, and the cost soars. The Tool Search mechanism of GPT-5.4 naturally solves this problem - there is no need to insert all tool definitions into the context, and they can be accessed on demand, and the token consumption is reduced by 47%. For OpenClaw running 24/7, this could mean saving tens or even hundreds of dollars in monthly API fees.
A jump in reasoning power—say goodbye to advanced autocompletion.
I used a low-end model to run OpenClaw before, but it got stuck when encountering slightly complex tasks. GPT-5.4 outperforms 83% of human experts in professional work tasks. Coupled with its comprehensive improvements in mathematics, programming, document processing and other fields, it means that OpenClaw can handle more complex and higher-value work scenarios. From helping you write code and fix bugs, to helping you do financial analysis and write investment memos, your imagination is instantly full.
In fact, OpenClaw's official documentation already provides access configuration solutions for GPT-5.4, and OpenAI's Responses API is also fully supported.
The first group of early adopters in the community should already be configuring it.
The starting point of a new era
Let's zoom out a little bit.
The release of GPT-5.4 is not only another refresh of model parameters, but also marks that the AI industry is undergoing a switch in the underlying logic:The leap from conversational AI to agent AI.
In the past few years, we have been discussing how smart AI is, what it can write, and what it can answer.
But starting from GPT-5.4, the topic has changed - what can AI do, what can it accomplish, and what can it replace.
When a model can independently control a computer and is more proficient than ordinary people, and can compete with industry experts in 83% of professional scenarios - this is not a quantitative change, this is a qualitative change.
Open source Agent frameworks like OpenClaw are the bridge that transforms this ability into actual productivity.
Combination of GPT-5.4+OpenClaw, let us clearly see for the first time the path of personal AI employees from concept to reality.
For some, the singularity has arrived.
Are you ready?
References:
https://chatgpt.com/codex